AI-Driven Medical Triage and Diagnostics: Lessons from a Special Forces Medic

By Jeremy Cleland — former U.S. Army Special Forces medic (18D) and AI engineer. M.S. in Artificial Intelligence, University of Michigan-Dearborn.

Updated 21 September 2026; originally published 09 May 2025. This revision reflects the TCCC Guidelines of 01 May 2026, the FDA's AI-enabled medical device list as of 29 June 2026, the 18 August 2026 FDA discussion paper on generative AI devices, and the state of the field as of September 2026.

Scope: this is an engineering and regulatory review, written for people who build software. It describes what published doctrine and public regulatory records say. It is not clinical guidance and not medical advice. The protocol values quoted below are here to show what a system would have to encode; anyone actually treating a casualty should be working from the current official guideline, within their own scope of practice, and not from this page.

Introduction: What Is AI-Driven Battlefield Triage?

AI-driven battlefield triage is software that sorts combat casualties by treatment and evacuation priority, scoring vital signs, imaging, or text against a protocol such as TCCC's MARCH/PAWS sequence. As of September 2026 the FDA-cleared examples are narrow: a three-vital-sign hemorrhage indicator and on-device ultrasound AI. Language-model triage assistants remain experimental everywhere.

That is the state of the art. The protocol all of it is built against is older and more interesting. MARCH is not a checklist. It is an ordering argument.

Massive hemorrhage, airway, respiration, circulation, head injury and hypothermia. The sequence is a claim about the clock: extremity arterial bleeding can kill in two to three minutes, an obstructed airway in four to six, a tension pneumothorax somewhat slower, and hypothermia slower still but with a coagulopathy penalty that compounds everything above it. The mnemonic does not tell you what matters. It tells you what matters first, given that you cannot attend to all of it and cannot know yet which problem the casualty actually has.

That is a triage of attention under uncertainty, and it is the property most AI triage systems lack. A model that emits a severity score has told you how worried to be. It has not told you what to do next, in what order, or what it deprioritized to get there. The interesting engineering question is not whether a model can predict mortality from vital signs — several can, and the literature is reasonably good on this. It is whether the output arrives in a structure a medic can act on inside the window the protocol assumes.

This post covers three things: what the current tactical combat casualty care protocol actually says as of 01 May 2026, what battlefield medical AI looks like in the field right now, and the evidence that argues against deploying it quickly. I have tried to keep the third part as long as the second, because the gap between the two is the story.

Consider a casualty presenting with a partially controlled thigh wound, a heart rate of 124, a systolic pressure of 86, and, after a blast, an inability to follow simple commands. That is an illustrative case, not a memory; the same numbers reappear in the reference implementation below. My claim to your attention here is protocol fluency plus production machine-learning work, not anecdote; section 9 sets out the specific places where that domain knowledge changes an engineering decision rather than decorating it.

1. MARCH/PAWS and What Changed in the 01 May 2026 TCCC Guidelines

The current tactical combat casualty care protocol is the TCCC Guidelines, 01 May 2026, issued by the Committee on Tactical Combat Casualty Care and summarized in Deaton, Montgomery and Butler, Journal of Special Operations Medicine 26(2):89. The committee chair is CAPT Travis Deaton and the co-chair is Harold Montgomery. The authoritative platform is Deployed Medicine, which has required a login since December 2023. Note that the Joint Trauma System CPG index still lists the 25 January 2024 revision, so the index is not the place to check currency.

Two structural changes are worth knowing before the table. TACEVAC guidance is now a separate document maintained by the Committee on En Route Combat Casualty Care, and the 2026 release adds Supplement A – Triage in TCCC, with refined Role 1 triage recommendations. If you are building triage software against TCCC, that supplement is now the governing text rather than an inference from the main guideline.

The table below summarizes what the published guideline states, for the purpose of showing software designers what the protocol commits them to. The imperative phrasing belongs to the doctrine, not to this post, and the summary is not a substitute for the guideline itself.

The most operationally significant change in the TCCC Guidelines, 01 May 2026 is tourniquet conversion: convert within 2 hours if the casualty is not in shock, the wound is closely monitorable, and it is not an amputation, and do not remove a tourniquet in place more than 6 hours without close monitoring and laboratory capability. That change is also the one that reallocates training time. Business Insider's reporting on how Ukraine is changing U.S. tourniquet training describes the operational driver: evacuation timelines long enough that a device applied correctly and left in place becomes its own injury. The same publication's reporting that the next war may leave wounded troops waiting weeks for evacuation is the planning assumption behind it.

One terminology note, because it affects what you search for and what you cite. The Joint Trauma Lexicon now distinguishes prolonged casualty care (PCC) from prolonged field care. PCC is Role 1 care delivered for extended periods when the tactical situation prevents prompt evacuation, and it applies to all Role 1 responders, medical and non-medical. Prolonged field care is a continuation of PCC by advanced trained personnel. PCC conceptually replaces the older lexicon definition, so generic use of "prolonged field care" is superseded. The governing document is the Prolonged Casualty Care Guidelines, 21 December 2021, with nursing interventions, wound care and splint management updated 8 July 2025.

2. The 2026 State of Battlefield Medical AI

2.1 ALICE vs FieldCare GPT: cloud or offline?

Two battlefield triage assistants reported in 2026 made opposite architectural choices: ALICE runs on Google Gemini in the cloud, and FieldCare GPT runs fully offline on a laptop. Both were described by Kelsey Baker for Business Insider on 7 September 2026, and comparing them is the clearest way to understand where the field is right now.

ALICE, also called the Tactical Aid Assistant, was built by Roman Kuziv, a Ukrainian trauma surgeon who commands the Medical Forces and Sustainment Group East in the Armed Forces of Ukraine. It runs on Google Gemini and is therefore cloud-dependent. The interaction is image-first: a soldier photographs an injury, answers follow-up questions, and receives the relevant TCCC protocol. It has separate modes for medics and non-medics, and voice prompts for users who are too distressed to type. It is in testing with Ukrainian medics at facilities near the front, with feedback flowing back to the developers. Kuziv describes it as protocol-oriented and explicitly not a replacement for a physician, and posted in August 2026 that the team was recruiting a software-as-a-medical-device regulatory lawyer.

FieldCare GPT was built by the University of Pittsburgh School of Medicine's Computational Pathology and AI Center of Excellence — not UC Davis, an error that circulates because senior technical director Nam Tran previously worked there. The team includes Tran, Hooman Rashidi, Ronald Poropatich, and Travis Polk of Pitt's Center for Military Medicine Research. It is a fully offline multi-agent retrieval-augmented chatbot for TCCC, running locally on a high-performance laptop with no internet connection. Its knowledge base is the 75th Ranger Regiment's Ranger Medic Handbook, the TCCC guidelines, and other combat trauma texts, and answers trace back to their source documents. It has been tested against 200 prompts related to battlefield injuries and demonstrated to military personnel, and it was shown at Exercise Northern Strike 2026-2 at the Alpena Combat Readiness Training Center. Validation data was presented at MHSRS 2026 but the figures are not public, so nobody outside the team can quote an accuracy number, including me. The name is a placeholder. It began after Ukrainian counterparts asked for fast, low-signature guidance that would not hallucinate.

The most instructive detail is what FieldCare GPT deliberately does not do. It has no image recognition at all. Tran's stated reasons are poor lighting, blood, clothing, dirt, and catastrophic injury. That is the best engineering judgment in the public record on this topic, and it is a direct rejection of the interaction model ALICE is built on. Neither team is wrong. They are optimizing for different environments.

2.2 Can AI run offline in a communications-denied environment?

The usual argument for on-device inference is privacy or latency. On a battlefield it is neither. Business Insider states the constraint plainly: cellphones, radios, and other powered electronics produce electromagnetic signatures that enemy forces can potentially detect and use for targeting. A laptop running local inference emits. A cloud-connected phone emits and transmits.

Denied, degraded, intermittent and limited-bandwidth conditions then cascade into everything else an ML team normally relies on. No cloud inference. No model updates in theater. No telemetry back. No remote monitoring. No over-the-air rollback of a model that turns out to be wrong. Weights, knowledge base, and guardrails all have to be resident before the mission starts, and whatever is wrong with them stays wrong for the duration.

This is also why cloud oncology and radiology systems are not the relevant comparison. Microsoft's HIMSS announcements — MedImageParse 3D, Rad-DINO, the Healthcare Agent Orchestrator, and a healthcare AI model leaderboard built with Stanford — are 2025 announcements of capable cloud models oriented toward oncology, radiology, and pathology, and none of them is a trauma or triage tool. That is the gap an offline tool addresses.

Chinese state media reports a parallel effort worth noting with appropriate caution. Science and Technology Daily reported on 16 June 2026 that Army Medical University's Xinqiao Hospital deployed two AI diagnostic screening systems to grassroots units: an X-ray system covering 20 abnormality classes at one to two seconds per plate, and an ultrasound system described as the first PLA-wide multi-disease offline system, covering 30 conditions and not dependent on a network. There is no independent verification of any of this, and the claim is diagnostic screening rather than triage. The offline design choice is the part that tracks with everything else in this section.

2.3 What can rugged edge hardware actually run?

The specification sheets for rugged edge compute are impressive and slightly misleading. Connect Tech's Sentry-X2 Mini, announced 12 May 2026, is a Jetson Orin NX 16GB system rated at 157 TOPS, MIL-STD-810H, DO-160H, MIL-STD-461G and MIL-STD-704F qualified, IP67 with D38999 connectors, TPM 2.0, rated from −20 °C to +70 °C, at 1.5 kg. Its larger sibling the Sentry-X2 uses the Jetson AGX Orin Industrial at 248 TOPS with 64 GB of ECC LPDDR5. Aitech's A230 Vortex is a fanless, IP66, cold-plate-cooled equivalent, and 7StarLake's NV200 is an Orin NX platform with no moving parts rated from −40 °C to +55 °C.

Here is the part that matters for a language model. TOPS is not LLM throughput. Those 157 and 248 TOPS figures are INT8 vision-inference numbers. Token generation is memory-bandwidth-bound, not compute-bound. The Orin NX's governing constraint is its 102.4 GB/s of memory bandwidth, which is well below a laptop GPU. That is very likely why FieldCare GPT is laptop-based, and it is consistent with Tran's comment that smaller devices cannot yet support such a sophisticated chatbot without significant performance compromises.

The quantization arithmetic is straightforward. GGUF-format models load memory-mapped with per-layer mixed quantization. For a 7B model: Q2_K is about 2.7 GB with significant degradation, Q3_K_M about 3.6 GB with noticeable loss, Q4_K_M about 4.9 GB and the sensible default, Q5_K_M 5.7 GB and near-FP16, Q8_0 8.0 GB and effectively indistinguishable from FP16, F16 14.0 GB. The working rule is that Q4_K_M needs roughly 0.6 to 0.7 GB per billion parameters. The token rates that follow are a memory-bandwidth illustration, measured on Llama 3-class 8B models at Q4_K_M. They are not a ranking of current models. On that class, throughput runs about 5–8 tokens per second on CPU alone with 64 GB of RAM, 10–15 on a modern desktop CPU with AVX2, 15–20 on an M2 MacBook Air, 30–40 on an M3 Pro, 40–60 on an RTX 3060, and 100–120 on an RTX 4090. A 70B model at Q4_K_M manages 15–20 on a 4090. Quantizing the KV cache with --cache-type-k q8_0 --cache-type-v q4_0 roughly halves its footprint with minimal quality loss, which matters when you are holding long protocol documents in context.

Vision on device is mature enough to run and not mature enough to trust. In llama.cpp, vision models ship as a pair: the quantized language model plus a separate multimodal projector file, and standard practice is to quantize the language model to Q4_K_M or Q8_0 while keeping the projector at F16 or Q8_0. Gemma 3 4B at Q4_K_XL is about 2.4 GB of language model plus 812 MB of projector, roughly 3.2 GB total, and runs on a MacBook. But in Simon Willison's hands-on notes on llama.cpp's vision support, Gemma 3 4B miscounted objects in a photograph. Small on-device vision-language models are runnable, not reliable, for trauma imagery — which is exactly the failure mode Tran cited.

Power closes the loop. Orin NX-class rugged platforms draw roughly 10 to 25 W and expect MIL-STD-704F-conditioned vehicle power. A high-performance inference laptop is a meaningful battery load in a dismounted prolonged-casualty-care scenario measured in days to weeks.

2.4 What battlefield medical AI is already FDA-cleared?

Most of the battlefield-relevant medical AI that is already cleared and shipping is handheld ultrasound rather than triage software. The cleared exceptions, covered in section 6, are a three-vital-sign hemorrhage indicator and a burn-assessment device — both narrow and both explicitly adjunctive. Exo announced in June 2025 the first FDA 510(k)-cleared AI for detecting pleural effusion and consolidation or atelectasis, bringing Exo Iris to 14 cleared AI indicators embedded in the device. Exo states that all of its commercialized AI runs directly on Iris and operates without internet.

Clarius reports six FDA-cleared models and 16 deployed AI technologies, per its 2025 year-end announcement, and on 2 June 2026 received clearance for an Ejection Fraction AI providing automated left-ventricular function assessment, positioned by the company for resuscitation guidance, undifferentiated shortness of breath, and trauma.

That is the realistic near-term picture: narrow, cleared, single-organ-system inference running on a handheld probe with no connectivity requirement. It is less exciting than a chatbot and considerably further along.

2.5 AEGIS MEDIC and Neolens: claims that are not yet products

AEGIS MEDIC, from Wilson Intelligence Services, is a useful case study in reading a pre-clearance claim carefully. The company describes hardware-agnostic middleware that turns a handheld ultrasound probe — it names Butterfly iQ, Clarius, and Philips Lumify as targets — into a guided eFAST exam, combining a YOLO model for probe guidance, a U-Net for pathology segmentation, and a rule-based engine that emits a TCCC T1–T4 category and a MEDEVAC 9-line. All of that is company self-description.

The verifiable status: the company states it has filed non-provisional patent application 19/688,914 on 27 May 2026, claiming priority to a February 2026 provisional. That filing is self-attested; U.S. applications are not published until roughly 18 months after priority, so there is no public file wrapper to check until around August 2027. The company says it is targeting a Class II 510(k) submission and says it is engaged with DIU and has submitted an MTEC bridge-loan application. It is not cleared, not submitted, not funded, and not clinically validated. Its own posting from 1 June 2026 describes a working demo covering four scenarios across 80 ultrasound frames with 51 automated tests passing, and lists clinical model training on real eFAST datasets and live probe integration as the next 90 days of work. As of that posting the models had not been trained on real clinical eFAST data and no live probe was integrated.

I am spelling this out at length because the gap between "has filed a patent application" and "has a cleared device" is the entire regulatory section below, and the language in this market collapses it constantly. A related pattern shows up in offline marketing claims generally: the Ukrainian startup Neolens markets a prolonged field care product page describing fully offline operation, while its founders told Militarnyi that Ukrainian users currently all use the web version, that offline requires an unreleased mobile app, and that the offline model is slightly smaller and gives narrower answers. Its verifiable deployment is retrieval over technical manuals for vehicle repair and weapons operation, not medicine.

2.6 Where military AI is actually working: medic training analytics

The most clearly successful military application of AI in this space is not clinical. At the Joint Special Operations Medical Training Center at Fort Bragg — the roughly nine-month Special Operations Combat Medic pipeline and the four-month Special Forces Medical Sergeant Course — leaders enter student performance data, including time to master each skill and the repetitions required, into what the course chief calls a calculator. Analytics then flag weak points across course iterations. Col. Ken Dwyer, the school's commander, told Business Insider in July 2026 that the constraint is time with the students before they go to the force. The worked example is tourniquets. Application was effectively the whole skill in the post-2001 wars, so where the numbers show a head start, hours can move toward tourniquet conversion, infection, and medication management. Business Insider frames that as a reallocation the data makes available, not one already banked. There is no model, no inference, and no clinical risk. The output is a schedule.

3. How MARCH Observations Map onto AI Model Features

The reason protocol fluency matters to a machine-learning problem is that it tells you which observations are load-bearing. A medic does not record everything; the assessment sequence is itself a feature-selection prior, arrived at through a long and expensive process. Mapping it onto model inputs is the cheapest domain knowledge available.

Note that the right-hand column is mostly trends, not point values. The single most common modeling error in this space is treating a casualty as a snapshot when the protocol treats them as a trajectory, reassessed after each intervention. TCCC's 2026 circulation guidance is explicit about reassessing after every unit of blood; a model that scores the initial vitals and stops has thrown away the signal the protocol considers most informative.

The compensatory reserve literature is the clearest example of trend-based measurement done properly: see the 2024 Army Warfighter Expeditionary Experiment field test and the 2026 validation of a wearable photoplethysmography sensor in simulated hemorrhage. The underlying device type was established by FDA De Novo DEN160020 in 2016.

4. A Reference Implementation: Encoding MARCH/PAWS in Code

What follows is an illustrative reference implementation, not a validated or tested system. Its purpose is to show how the protocol's structure — ordered assessment, explicit thresholds, reassessment, and a human decision point — maps onto code. It has not been trained on clinical data, has not been evaluated, and should not be used for anything. The model objects are unfitted; calling predict on them would raise. Read it as a schema for encoding a protocol, which is the part that is actually transferable.

"""
battlefield_triage_ai.py

ILLUSTRATIVE REFERENCE IMPLEMENTATION — NOT A VALIDATED SYSTEM.

Demonstrates how the MARCH/PAWS assessment sequence maps onto a feature
schema, threshold logic, and an explainability surface. Protocol
constants below are from the TCCC Guidelines, 01 May 2026. Shock-index
cutpoints are illustrative and are not TCCC thresholds. No model here
has been trained on clinical data; none of this has been evaluated
for clinical use.
"""

import pandas as pd
import numpy as np
from datetime import datetime
from sklearn.ensemble import RandomForestClassifier
from sklearn.preprocessing import StandardScaler
import shap
import logging

logging.basicConfig(
    level=logging.INFO,
    format='%(asctime)s [%(levelname)s] %(message)s',
    datefmt='%Y-%m-%d %H:%M:%S'
)
logger = logging.getLogger(__name__)

# TCCC Guidelines, 01 May 2026.
TQ_CONVERT_WITHIN_MINUTES = 120
TQ_DO_NOT_REMOVE_AFTER_MINUTES = 360

# Illustrative transfusion-literature cutpoints. Not TCCC thresholds.
ILLUSTRATIVE_SHOCK_INDEX = 1.0
ILLUSTRATIVE_SHOCK_INDEX_URGENT = 1.5


class BattlefieldTriageAI:
    """
    MARCH/PAWS-structured triage scaffold with explainable output.

    The design intent is that the system augments and never replaces
    the medic's judgment. The final decision rests with the human
    on scene.
    """

    def __init__(self):
        self.model = RandomForestClassifier(
            n_estimators=100,
            class_weight='balanced',
            random_state=42,
            n_jobs=-1
        )
        self.scaler = StandardScaler()
        self.internal_hemorrhage_model = RandomForestClassifier(
            n_estimators=50,
            random_state=42
        )
        logger.info("Triage scaffold initialized (untrained)")

    def calculate_shock_index(self, heart_rate: float, systolic_bp: float) -> float:
        """
        Shock Index = HR / SBP.

        Association with massive transfusion and mortality in combat
        casualties is established in the DoD Trauma Registry literature:
          - Shock index and pulse pressure as predictors of massive
            transfusion and death (DoDTR, n=15,540), PMID 34449863
          - Shock index as a predictor of massive transfusion and
            emergency surgery on the modern battlefield
            (DoDTR 2008-2016, n=4,008), PMID 32683051
          - Systematic review and meta-analysis (n=670,728), PMC9985849

        The thresholds below are illustrative cutpoints, not a protocol.
        """
        if systolic_bp <= 0:
            logger.warning("Invalid SBP; returning saturated shock index")
            return 2.0
        return heart_rate / systolic_bp

    def assess_hemorrhage_risk(self, vitals: dict) -> np.ndarray:
        """
        Encode indirect signs of ongoing or occult hemorrhage:
          - narrowing pulse pressure
          - elevated shock index
          - mechanism suggesting high-energy transfer
        """
        features = []

        pulse_pressure = vitals['systolic_bp'] - vitals['diastolic_bp']
        features.append(1 if pulse_pressure < 30 else 0)

        shock_index = self.calculate_shock_index(
            vitals['heart_rate'],
            vitals['systolic_bp']
        )
        features.append(1 if shock_index > ILLUSTRATIVE_SHOCK_INDEX else 0)

        moi_risk = vitals.get('blast_injury', 0)
        features.append(moi_risk)

        return np.array(features)

    def extract_march_features(self, patient_data: dict) -> dict:
        """
        Convert an assessment into a feature vector ordered by the
        MARCH/PAWS sequence, so that the schema is auditable against
        the protocol it claims to follow.
        """
        shock_index = self.calculate_shock_index(
            patient_data['heart_rate'],
            patient_data['systolic_bp']
        )
        pulse_pressure = patient_data['systolic_bp'] - patient_data['diastolic_bp']
        map_value = (patient_data['systolic_bp'] + 2 * patient_data['diastolic_bp']) / 3

        features = {
            # M - Massive hemorrhage
            'shock_index': shock_index,
            'pulse_pressure': pulse_pressure,
            'tourniquet_applied': patient_data.get('tourniquet_applied', 0),
            'tourniquet_minutes': patient_data.get('tourniquet_minutes', 0),
            'in_shock': patient_data.get('in_shock'),
            'wound_monitorable': patient_data.get('wound_monitorable'),
            'amputation': patient_data.get('amputation'),
            'external_bleeding_controlled': patient_data.get('bleeding_controlled', 0),

            # A - Airway
            'spo2': patient_data.get('spo2', 98),
            'airway_patent': patient_data.get('airway_patent', 1),
            'stridor_present': patient_data.get('stridor', 0),

            # R - Respiration
            'respiratory_rate': patient_data['respiratory_rate'],
            'breath_sounds_equal': patient_data.get('equal_breath_sounds', 1),
            'ndc_attempts': patient_data.get('ndc_attempts', 0),

            # C - Circulation
            'mean_arterial_pressure': map_value,
            'radial_pulse_palpable': patient_data.get('radial_pulse', 1),
            'txa_administered': patient_data.get('txa_given', 0),
            'units_transfused': patient_data.get('units_transfused', 0),
            'calcium_given': patient_data.get('calcium_given', 0),

            # H - Head injury / hypothermia
            'follows_commands': patient_data.get('follows_commands', 1),
            'minutes_since_injury': patient_data.get('injury_time_minutes', 0),
            'pupil_response': patient_data.get('pupils_equal', 1),
            'core_temperature': patient_data.get('temperature', 37.0),

            # Mechanism
            'mechanism_blast': patient_data.get('blast_injury', 0),
        }

        features['hemorrhage_risk_score'] = self.assess_hemorrhage_risk(patient_data).sum()

        return features

    def predict_priority(self, patient_vitals: dict) -> dict:
        """
        Produce a triage category with a confidence estimate and a
        structured MARCH report. Requires fitted models.
        """
        features = self.extract_march_features(patient_vitals)
        features_df = pd.DataFrame([features])

        features_scaled = self.scaler.transform(features_df)

        priority = self.model.predict(features_scaled)[0]
        priority_proba = self.model.predict_proba(features_scaled)[0]

        hemorrhage_features = self.assess_hemorrhage_risk(patient_vitals)
        internal_risk = self.internal_hemorrhage_model.predict_proba(
            hemorrhage_features.reshape(1, -1)
        )[0][1]

        assessment = self._generate_march_report(features, priority, patient_vitals)

        return {
            'triage_category': self._map_priority_to_category(priority),
            'priority_score': priority,
            'confidence': float(max(priority_proba)),
            'internal_hemorrhage_risk': float(internal_risk),
            'march_assessment': assessment,
            'recommended_actions': self._get_immediate_actions(features, priority),
            'evacuation_priority': self._determine_evac_priority(features, priority)
        }

    def _generate_march_report(self, features: dict, priority: int,
                               raw_vitals: dict) -> list:
        """
        Emit the assessment in MARCH order, which is how it would be
        reported up the chain. Ordering is part of the output contract.
        """
        report = []

        # M - Massive hemorrhage
        if features['shock_index'] > ILLUSTRATIVE_SHOCK_INDEX:
            report.append(
                "M: CRITICAL - Illustrative cutpoint, shock index above 1.0 "
                "(not a TCCC threshold); hemorrhage likely"
            )
        elif features['pulse_pressure'] < 30:
            report.append("M: WARNING - Narrow pulse pressure, occult bleeding possible")
        elif not features['external_bleeding_controlled']:
            report.append("M: URGENT - External hemorrhage not yet controlled")
        else:
            report.append("M: STABLE - No active hemorrhage detected")

        # M - Tourniquet reassessment (TCCC 01 May 2026).
        # More than 6 hours, then the 2-to-6-hour band, then under 2 hours.
        tq_minutes = features['tourniquet_minutes']
        if features['tourniquet_applied'] and tq_minutes > TQ_DO_NOT_REMOVE_AFTER_MINUTES:
            report.append(
                f"M: TQ in place {tq_minutes} min - more than 6 h; "
                "do not remove without close monitoring and lab capability"
            )
        elif features['tourniquet_applied'] and tq_minutes >= TQ_CONVERT_WITHIN_MINUTES:
            report.append(
                f"M: TQ in place {tq_minutes} min - conversion window passed; "
                "conversion requires Combat Medic/Paramedic/Provider direction"
            )
        elif features['tourniquet_applied']:
            conditions_met = (
                features['in_shock'] == 0
                and features['wound_monitorable'] == 1
                and features['amputation'] == 0
            )
            condition_status = "conditions met" if conditions_met else "conditions not met"
            report.append(
                f"M: TQ in place {tq_minutes} min - under 2 h; convert only if "
                "not in shock, wound closely monitorable, and not an amputation "
                f"({condition_status})"
            )

        # A - Airway
        if features['spo2'] < 90:
            report.append("A: CRITICAL - Severe hypoxia, immediate airway intervention")
        elif features['stridor_present']:
            report.append("A: URGENT - Stridor present, airway compromise developing")
        elif not features['airway_patent']:
            report.append("A: WARNING - Airway requires continuous monitoring")
        else:
            report.append("A: CLEAR - Airway patent and protected")

        # R - Respiration
        rr = features['respiratory_rate']
        if rr > 29 or rr < 10:
            report.append(f"R: CRITICAL - RR {rr}, consider NDC if indicated")
        elif not features['breath_sounds_equal']:
            report.append("R: WARNING - Unequal breath sounds, monitor for pneumothorax")
        else:
            report.append("R: ADEQUATE - Normal respiratory pattern")

        # C - Circulation
        map_val = features['mean_arterial_pressure']
        if not features['radial_pulse_palpable']:
            report.append("C: CRITICAL - No palpable radial pulse")
        elif map_val < 65:
            report.append(f"C: CRITICAL - MAP {map_val:.0f}, damage control resuscitation")
        else:
            report.append("C: STABLE - Hemodynamically stable")

        if features['units_transfused'] > 0 and not features['calcium_given']:
            report.append("C: GAP - Calcium not given after first transfused product")

        # H - Head injury / hypothermia
        temp = features['core_temperature']
        if not features['follows_commands'] and features['minutes_since_injury'] > 10:
            report.append(
                "H: Unable to follow commands beyond 10 min post-injury - "
                "treat as moderate/severe TBI; target SpO2 at or above 92%, "
                "SBP above 100 mmHg; neurosurgery within 5 h"
            )
        if temp < 36:
            report.append(f"H: HYPOTHERMIC - Core temp {temp} C, begin active warming")

        return report

    def _map_priority_to_category(self, priority: int) -> str:
        """Map numeric priority to NATO triage categories."""
        categories = {
            1: "IMMEDIATE (T1)",
            2: "DELAYED (T2)",
            3: "MINIMAL (T3)",
            4: "EXPECTANT (T4)"
        }
        return categories.get(priority, "IMMEDIATE (T1)")

    def _get_immediate_actions(self, features: dict, priority: int) -> list:
        """
        Surface protocol-derived prompts. These are reminders against a
        published guideline, not recommendations, and they do not
        substitute for the medic's assessment.
        """
        actions = []

        if features['shock_index'] > ILLUSTRATIVE_SHOCK_INDEX and not features['tourniquet_applied']:
            actions.append("Reassess for a controllable hemorrhage source")
        if features['hemorrhage_risk_score'] > 1 and not features['txa_administered']:
            actions.append(
                "TXA 2 g slow IV/IO push - must be within 3 h of injury"
            )

        if features['spo2'] < 90:
            actions.append("Airway intervention; surgical cricothyroidotomy if obstructed")
        elif features['stridor_present']:
            actions.append("Position for rapid airway intervention")

        if features['respiratory_rate'] > 29 and features['ndc_attempts'] < 2:
            actions.append(
                "Assess for tension pneumothorax - NDC with 14 ga or 10 ga "
                "3.25 in unit; max 2 attempts, second at the other site"
            )

        if features['mean_arterial_pressure'] < 65 or not features['radial_pulse_palpable']:
            actions.append(
                "Transfuse in TCCC preference order, beginning with cold-stored "
                "low-titer O whole blood; reassess after each unit"
            )
            if not features['calcium_given']:
                actions.append(
                    "1 g calcium IV/IO (30 mL 10% gluconate or 10 mL 10% chloride) "
                    "after the first transfused product"
                )

        if features['core_temperature'] < 36:
            actions.append(
                "Active heating blanket to anterior torso and axillae, then "
                "impermeable enclosure; warm fluids at 38 C"
            )

        return actions

    def _determine_evac_priority(self, features: dict, priority: int) -> str:
        """
        Evacuation precedence. Reference: ATP 4-02.2, Medical Evacuation.
        """
        if priority == 1 or features['shock_index'] > ILLUSTRATIVE_SHOCK_INDEX_URGENT:
            return "URGENT - Evacuate within 1 hour"
        elif priority == 2 or features['hemorrhage_risk_score'] > 1:
            return "PRIORITY - Evacuate within 4 hours"
        elif priority == 3:
            return "ROUTINE - Evacuate within 24 hours"
        else:
            return "CONVENIENCE - Evacuate when possible"

    def explain_decision(self, patient_vitals: dict) -> dict:
        """
        Expose the drivers behind the score. Goddard et al. found that
        providing information rather than a recommendation, and attaching
        confidence to output, mitigate automation bias. That is the
        design goal here.
        """
        features = self.extract_march_features(patient_vitals)
        features_df = pd.DataFrame([features])

        explainer = shap.TreeExplainer(self.model)
        shap_values = explainer.shap_values(self.scaler.transform(features_df))

        if isinstance(shap_values, list):
            shap_values = shap_values[1]

        feature_importance = dict(zip(features_df.columns, shap_values[0]))

        top_factors = sorted(
            feature_importance.items(),
            key=lambda x: abs(x[1]),
            reverse=True
        )[:5]

        explanation = {
            'primary_drivers': [],
            'confidence_factors': [],
            'risk_indicators': []
        }

        for feature, importance in top_factors:
            label = feature.replace('_', ' ').title()
            if importance > 0:
                explanation['risk_indicators'].append(
                    f"{label}: increases priority (impact: {importance:.3f})"
                )
            else:
                explanation['confidence_factors'].append(
                    f"{label}: decreases priority (impact: {importance:.3f})"
                )

        if features['shock_index'] > ILLUSTRATIVE_SHOCK_INDEX:
            explanation['primary_drivers'].append(
                "Illustrative cutpoint: shock index above 1.0 is associated "
                "with massive transfusion, and is not a TCCC threshold"
            )
        if features['spo2'] < 90:
            explanation['primary_drivers'].append(
                "Hypoxia below 90% requires immediate airway attention"
            )

        return explanation


# Illustrative input only. Models above are unfitted; this block shows the
# expected shape of an assessment, not a working demonstration.
if __name__ == "__main__":
    logger.info("Illustrative reference implementation - not for clinical use")
    scaffold = BattlefieldTriageAI()

    illustrative_casualty = {
        'heart_rate': 124,
        'systolic_bp': 86,
        'diastolic_bp': 52,
        'respiratory_rate': 28,
        'spo2': 89,
        'temperature': 35.5,
        'follows_commands': 0,
        'blast_injury': 1,
        'tourniquet_applied': 1,
        'tourniquet_minutes': 140,
        'txa_given': 0,
        'units_transfused': 1,
        'calcium_given': 0,
        'equal_breath_sounds': 0,
        'injury_time_minutes': 15
    }

    features = scaffold.extract_march_features(illustrative_casualty)
    report = scaffold._generate_march_report(features, 1, illustrative_casualty)
    actions = scaffold._get_immediate_actions(features, 1)

    print("\n" + "=" * 50)
    print("MARCH/PAWS STRUCTURED ASSESSMENT (ILLUSTRATIVE)")
    print("=" * 50)
    print(f"Time: {datetime.now().strftime('%H:%M:%S')}")

    print("\nMARCH assessment:")
    for item in report:
        print(f"  • {item}")

    print("\nProtocol prompts:")
    for action in actions:
        print(f"  → {action}")

Three properties of that scaffold are the ones worth carrying into a real system. The output is ordered by the protocol rather than by model confidence, so a reader can check it against the guideline. The protocol values the 01 May 2026 guideline actually states — tourniquet time, the conversion conditions, TXA within three hours, calcium after the first unit, the ten-minute command window — are named constants. Shock-index cutpoints are labeled illustrative in the report. An index above 1.0 is a transfusion-literature association, not a TCCC threshold. And the action list is phrased as protocol reminders rather than recommendations — a distinction supported by the automation-bias literature in section 7.

5. Beyond Trauma: AI for Austere and Prolonged Casualty Care

Trauma is the part people picture, but a Role 1 provider in a remote area spends most of their time on infectious disease, chronic complaints, and the general practice of medicine without a laboratory. That is a different AI problem: less about time pressure, more about differential diagnosis with an incomplete history and no confirmatory testing.

Consider an illustrative case. A child presents with a several-day fever, a rash, and lethargy, in a region where dengue, chikungunya, rickettsial disease, and malaria are all plausible. There is no laboratory, the history is coming through an interpreter, and evacuation is not free. What would actually help is not a diagnosis. It is a structured differential with region-appropriate priors, an explicit list of the findings that would discriminate between the candidates, and a clear statement of the red flags that convert this from a clinic visit into an evacuation.

This is where a retrieval-based system like FieldCare GPT is architecturally better suited than an image classifier. The task is to surface the relevant section of a reference text, with its provenance attached, faster than a provider could find it themselves — not to replace their judgment about which text applies. It also sidesteps the problem Tran identified: a photograph of a rash in poor light on unfamiliar skin tone is exactly where small vision models fail.

6. The Regulatory Picture: FDA, PCCP, and the EU AI Act

What has the FDA actually authorized for medical AI?

The FDA's list of AI-enabled medical devices held 1,614 devices as of 29 June 2026. Checked again on 21 September 2026, that was still the newest decision date on the published list. Of those, 1,553 (96.2%) came through 510(k), 40 (2.5%) through De Novo, and 21 (1.3%) through PMA. 76.2% are radiology. Growth has been steep: 41 devices between 1995 and 2015, 162 in 2022, 235 in 2024, 335 in 2025, and 181 in the first half of 2026 alone. The FDA notes the list is not comprehensive and that inclusion reflects AI terminology appearing in marketing authorization summaries rather than any certification of performance.

According to Congressional Research Service brief IF13245, the FDA "does not appear to have authorized for marketing any generative AI (genAI)-enabled device." On 18 August 2026 the agency published a discussion paper, Considerations for the Regulation of Generative AI-Enabled Medical Devices, docket FDA-2026-N-7874, with comments due 19 October 2026. It is not guidance, and it authorizes nothing. It is the document in which FDA asks how a language-model device, a foundation-model update, and an agentic system would be evaluated. That is the live regulatory question for anyone building a language-model battlefield assistant. In March 2026 the agency granted Breakthrough Device designation to a patient-facing clinical generative AI application from RecovryAI, which is a designation, not an authorization. An LLM-based battlefield medical assistant sits in a regulatory category with zero authorized devices.

What did the FDA grant DeepView in May 2026?

On 21 May 2026 the FDA granted De Novo classification DEN250028 to SpectralMD for the DeepView System, a software-aided adjunctive diagnostic device for skin wound assessment indicated for burn care. The company announced it on 26 May 2026. It is a Class II device under 21 CFR 878.1840, product code SHY, cleared for burn centers, trauma centers and emergency departments, using multispectral imaging and AI to predict burn healing potential in roughly 20 to 25 seconds. It was funded in part by BARDA contract 75A50123C00049. This was the real regulatory event of 2026. Note the framing: adjunctive, for healthcare providers, for burns. Not battlefield triage.

What is APPRAISE-HRI?

APPRAISE-HRI is the military hemorrhage-risk software people usually have in mind when they say battlefield AI is already fielded. Its record is 510(k) K233249, with a decision date of 5 April 2024, announced by USAMRDC on 11 April. It was cleared, not approved — 510(k) and De Novo produce clearances, only PMA produces approvals, and the two words are not interchangeable no matter how the press release phrases it. It is also not the first of its kind: the device type was created by De Novo DEN200022, granted to Fifth Eye on 1 March 2021, and APPRAISE-HRI cleared against that predicate. The defensible narrow claim is that it was the first AI-enabled decision-support software cleared under the military-specific "Adjunctive Hemodynamic Indicator With Decision Point for Military Use" classification.

What it actually does is considerably more modest than the coverage suggested. The applicant is the Surgeon General of the Department of the Army; it was developed by USAMRDC's BHSAI. It is an Android application that connects over Bluetooth to a ZOLL Propaq M monitor. Its inputs are heart rate, systolic pressure and diastolic pressure only. Its output is a Hemorrhage Risk Index at Levels I through III, produced by an AI-based linear regression model. Validation was a prospective-retrospective clinical utility study on n=5,895 trauma registry records, with likelihood ratios of 0.18 for Level I, 0.70 for Level II, and 6.88 for Level III. The indication states explicitly that it is not intended to diagnose or direct treatment. The Army holds three U.S. patents and its technology transfer office is seeking licensees, so this is a transition-seeking government asset rather than a fielded commercial product.

Can a 30-day model refresh coexist with FDA validation?

Here is the structural problem, and it is the strongest argument in this post.

The Artificial Intelligence Strategy for the Department of War, released January 2026, sets a goal of being "AI-first" and issues some unusually hard procurement directives. Vendors must be able to deploy the latest models within 30 days of public release, described in the document as a primary procurement criterion for future model acquisition. Model-objectivity benchmarks become a procurement criterion within 90 days. A monthly Barrier Removal Board under the Under Secretary for Research and Engineering has authority to waive non-statutory blockers, alongside rapid authority-to-operate reciprocity. This follows the realignment of the Chief Digital and AI Office under USD(R&E) in August 2025, which the Congressional Research Service characterized as an effective demotion and which the department framed as the way to accelerate AI capability delivery — the memo directing it insists that "all actions must accelerate, not impede, our ability to field AI capabilities."

The FDA's software-as-a-medical-device regime wants the opposite: a frozen, validated model, with any changes specified in advance. The mechanism designed to reconcile these is the Predetermined Change Control Plan, set out in the FDA's final guidance of August 2025, which requires a description of planned modifications, a modification protocol, and an impact assessment. It pre-authorizes a bounded set of changes; it does not support continuous learning.

And here is the reality check: neither APPRAISE-HRI nor DeepView has an authorized PCCP. Both FDA records read "Predetermined Change Control Plan Authorized: No." The mechanism exists. In military trauma AI, it is not being used. A 30-day model refresh cadence and a frozen validated model are not reconcilable by good intentions, and at present nobody has demonstrated the paperwork path.

When does the EU AI Act apply to medical devices?

The EU AI Act timeline was rewritten this year. Regulation (EU) 2026/1744, the Digital Omnibus on AI, passed the Parliament plenary on 16 June 2026 by 423 to 57 with 174 abstentions, was published on 24 July and entered into force on 27 July. The general date of application remains 2 August 2026, and the Article 50 transparency obligations were not deferred. But high-risk standalone systems under Annex III moved from 2 August 2026 to 2 December 2027, and high-risk systems embedded in regulated products under Annex I — which includes medical devices — moved from 2 August 2027 to 2 August 2028. That last date is the operative one for AI medical devices, and the new dates are fixed and unconditional; the proposed standards-availability trigger was abandoned.

7. Does AI Triage Work? What the Evidence Shows

This is the section that should determine how fast any of this gets fielded, and it is not encouraging.

How often do language models undertriage emergencies?

The most directly relevant study published this year is an evaluation of ChatGPT Health, OpenAI's consumer health tool launched in January 2026. The authors ran 60 clinician-authored vignettes across 21 clinical domains and 16 factorial conditions, producing 960 responses.

In that Nature Medicine 2026 evaluation, ChatGPT Health undertriaged 51.6% of gold-standard emergencies — 33 of 64 — routing them to evaluation in 24 to 48 hours instead of the emergency department. Accuracy followed an inverted U: 35.2% on nonurgent cases, 93.0% on semi-urgent, 76.9% on urgent, and 48.4% on emergencies. It was worst precisely at the extremes. The specific failures were diabetic ketoacidosis, impending respiratory failure with rising pCO₂, and metabolic acidosis, while classical emergencies like stroke and anaphylaxis were handled correctly.

The finding that should stop a battlefield deployment cold is the anchoring result. When family or friends minimized the symptoms, triage shifted significantly in edge cases, with an odds ratio of 11.7 (95% CI 3.7–36.6), mostly toward less urgent. The authors concluded that these findings raise safety concerns warranting prospective validation before consumer-scale deployment.

Now map that onto the use case. A panicked soldier describing a buddy's wound to an application is a minimizing, unreliable narrator under acute stress. That is not the edge case for a battlefield triage assistant. That is the default case.

The broader triage literature agrees. A study of 39,375 real emergency department patients at AHEPA University General Hospital compared seven language models against 25 emergency physicians' Emergency Severity Index assignments; the best performers, DeepSeek and Claude Sonnet 4, reached a weighted kappa of about 0.467 and raw accuracy of 61.7%, while GPT-5 Instant managed kappa 0.176. None achieved strong agreement. A May 2026 preprint testing 16 models on real-world ED triage found quadratic-weighted kappas from 0.18 to 0.75, expected calibration error from 0.099 to 0.355 with systematic overconfidence, and substantial output variability in 23% of cases — non-deterministic behavior on clinically relevant decisions. Its conclusion is worth internalizing: structured prompting mattered more than model architecture or size.

Does AI decision support cause automation bias?

Two older findings frame the human side of a triage display.

Goddard, Roudsari and Wyatt, JAMIA 2012, screened 13,821 papers and included 74. Their headline result: erroneous clinical decision support advice increased incorrect decisions by 26%. The mediators they identified are all present on a battlefield — workload, task complexity, time constraint. But the mitigators are the more useful part, because they are design instructions: emphasize user accountability, position advice carefully, attach updated confidence levels to output, and provide information rather than a recommendation. That last one is the empirical basis for the explainability argument this post has been making, and it is why the reference implementation above emits protocol reminders rather than a verdict. Lyell and Coiera, reviewing 890 papers across 1983–2015, added that automation bias occurs in single tasks, more in diagnosis than monitoring, and scales with cognitive load.

Deskilling — whether exposure to an assistive tool decays the unassisted skill — is a different argument, and it has one home. Budzyń et al. reported that unaided adenoma detection fell after centres adopted AI polyp detection. The argument from that finding belongs in the tourniquet-conversion essay, which is also where the later endoscopy evidence belongs. This survey does not re-try it. The design constraint that belongs here is the one Goddard already states.

Does FDA clearance mean a device improves patient outcomes?

Only 3 of 1,357 FDA-cleared AI/ML medical devices (0.2%) have been evaluated for patient-centered outcomes such as mortality, morbidity or readmissions, according to a systematic analysis in PLOS Digital Health covering clearances through 5 December 2025. The same analysis found 34 (2.5%) linked to registered prospective trials, 12 (0.9%) with posted results, and 12 (0.9%) with peer-reviewed publications. 62% of the designs were observational. The authors' conclusion: "Regulatory approval has outpaced clinical validation... Readiness should no longer be defined by FDA clearance alone."

A systematic review specific to AI triage of the injured in war conditions, screening 189 studies from 2015–2025 and including 17, reports genuinely strong numbers for narrow tasks: AUC 0.98–0.987 for vital-sign models predicting massive transfusion, sensitivity up to 94% for shock and hypovolemia, AUC 0.81 for life-saving intervention, mIoU 0.73 for shrapnel detection with YOLO and U-Net architectures, and 60–64% for language models on START. Its conclusion is the sentence to remember: "Most studies rely on laboratory or simulated datasets, highlighting the need for real-world field validation."

How should battlefield medical AI be positioned?

The best available statement of how these tools should be positioned comes from Stacy Shackelford, MD, a professor of emergency medicine and surgery and the former director of the military's Joint Trauma System, speaking to Business Insider about these specific systems. She calls them "decision assist" — a way to augment memory or help with difficult choices. She cautions that outcomes might become more uncertain when tools push inexperienced users beyond their training, while noting that the alternative for a non-medic is often being told "Okay, you're on your own."

On the tourniquet question specifically, Business Insider records her point that deciding whether one can safely come off requires judgment about the wound and patient condition that is difficult to teach in basic first-aid training, let alone through an app. On the other half she is quoted directly: "even if you focused on just taking off tourniquets in Ukraine, that would make a huge impact on their system." Both halves of that are true simultaneously, which is a better description of the state of the field than anything a vendor has said. Kuziv and Tran both stressed that their applications are not meant to replace medical personnel or training.

The scholarly version is Mathais, Cungi, Lamblin, Dubourg, Bordes and Boussen, Journal of Trauma and Acute Care Surgery 2026;101:S177–S187, which frames AI for large-scale combat triage explicitly as something other than a replacement for clinical judgement. They identify three functions — extending caregiver perception, sustaining cognition under pressure, and enabling anticipatory decision-making — and sort capabilities by horizon: wearable physiological sensors, early warning systems, digital casualty documentation and unmanned platforms are near-term deployable; multimodal fusion, predictive support, augmented-reality guidance and partial automation are mid-term; digital twins remain research. Their named constraints are the ones this post has walked through: DDIL environments, data quality, cognitive and ergonomic risk, automation bias, survivability on a transparent battlefield, and governance. A parallel review in the Romanian Journal of Military Medicine, covering 41 studies from 2000 to 2026, reaches the same place: poor data quality, algorithmic bias, and limited validation in real operational settings.

8. The Data Problem: Why the DoD Trauma Registry Limits Model Training

Every model in this domain is downstream of a registry that was not built for it.

The National Academies' A National Trauma Care System reports that among more than 4,000 Department of Defense Trauma Registry casualties from 2007 to 2010, 87% had no documentation of mechanism or location of injury, prehospital vital signs, or prehospital interventions. A separate analysis covering 2002 to 2009 found prehospital data in only 18.6% of Operation Enduring Freedom records and 25.4% of Operation Iraqi Freedom records. The Prehospital Trauma Registry was fielded in 2013; as of November 2015 roughly 750 patient records had been entered.

The registry's own limitations are documented by the people who run it. Military Review noted in January 2026 that DoDTR data is neither real-time nor standardized, requires manual abstraction, and suffered from a lack of commonality in tools, data formats, and entry methods. The DoDTR data dictionary states that registries are not primary research tools and are not designed to answer any specific scientific question, that missing data can present a greater concern, and that the registry includes only a selective sample. Injury coding accuracy is particularly poor in the sickest patients. DoDTR does not link to the civilian prehospital or hospital registries.

There is also a structural break that will quietly poison any model trained across it. Prehospital documentation methodology changed in 2007 and again in 2012. Before 2012, Role 1a, 1b and 1c care each had separate trauma records; afterward it was folded into the record for the first surgical-capable facility. A model trained across that boundary is learning a schema change as if it were a clinical signal.

Which brings up the substitution everyone makes. MIMIC-IV is excellent data and it is the wrong data. It is adult ICU and emergency department data from a single U.S. academic centre. Combat casualties are young, overwhelmingly male, with blast and penetrating mechanisms, evacuation timelines measured in hours to weeks rather than minutes, and resource constraints with no civilian analogue. A model validated on MIMIC-IV has been validated for a different population, a different injury distribution, and a different care pathway.

9. What Does Protocol Fluency Contribute to the Engineering?

None of the above requires having been a medic. But there are a few places where the domain knowledge changes an engineering decision rather than just decorating it.

Knowing what the protocol asserts, not just what it says. MARCH's ordering is a mortality-time-constant argument. That means the ordering is itself a feature of the output contract, and a system that reorders by model confidence has silently discarded clinical reasoning that was expensive to acquire. This is also why deviation is a skill: knowing why the sequence is ordered that way is what lets you recognize the case where it is wrong.

Knowing what gets documented and what does not. The 87% missingness figure above is not a data-quality accident. It reflects that documentation happens after the casualty is stable, from memory, by someone who has been awake for a long time. If your feature depends on a prehospital intervention timestamp, you should assume it is absent, wrong, or reconstructed.

Knowing the failure mode of the user, not just the model. A medic under load will not read a confidence interval. They will read the first line. Goddard's finding that providing information rather than a recommendation reduces automation bias is a design constraint, and satisfying it means the first line has to be the observation, not the verdict.

Being willing to say the system should refuse. The honest answer for a lot of battlefield cases is that the model does not have enough information and should say so. That is easier to argue for when you have done the job it is being built to assist. I have worked that argument out at length in why the Army's training analytics are a bet against the Department's AI strategy, and the fuller background is in my 18D and AI engineering bio.

10. Where Battlefield Medical AI Stands in September 2026

The state of the field as of September 2026, stated plainly:

Narrow, single-purpose, on-device imaging AI is cleared and shipping — Exo's 14 indicators, Clarius's six models. Physiological early-warning software is cleared and modest in what it claims: APPRAISE-HRI reads three vital signs and is explicitly not intended to diagnose or direct treatment. Burn assessment reached the market through De Novo in May 2026. Those are real capabilities with real regulatory records.

Language-model assistants are experimental everywhere. FieldCare GPT is offline, promising, and has no public performance data. ALICE is in the hands of medics near the front and depends on a cloud connection that the electromagnetic environment argues against. No generative AI device has been authorized by the FDA. The August 2026 discussion paper asks how one would be evaluated; it is not an authorization. The published triage evaluations show undertriage of the sickest patients and severe susceptibility to a minimizing narrator, which is the exact input a battlefield tool would receive.

The Department of War wants 30-day model refresh as a procurement criterion. The FDA wants frozen, validated models. Neither of the two relevant cleared devices has the mechanism that reconciles them. Meanwhile the DARPA Triage Challenge has not concluded. The public events page still shows a 30 September 2026 data deadline, systems finals 5–13 November, and awards on 13 November. The finals data rules open the data-submission window through 10 October, and the data-competition page lists awards on 14 November. The most recent completed event was Event 2, which concluded on 4 October 2025 at Guardian Centers. DARPA reported that DART and RoboScout scored highest overall in the systems competition and that MSAI excelled in the data competition; because only self-funded teams were eligible for prize money, Coordinated Robotics took $300,000 and UAS-DTU $150,000.

The case for spending those reclaimed hours on unassisted conversion, and the deskilling evidence with its limits, is in the tourniquet essay. What this survey can say on its own is Shackelford's framing: build the thing that augments memory and surfaces the protocol, attach the provenance, and let the medic decide. The evidence supports that. It does not yet support anything more.

References

1. Deaton TG, Montgomery HR, Butler FK. "Tactical Combat Casualty Care (TCCC) Guidelines: 1 May 2026 Updates." J Spec Oper Med. 26(2):89.
2. Deployed Medicine (canonical TCCC platform; login required).
3. Joint Trauma System, Committee on Tactical Combat Casualty Care.
4. Joint Trauma System, Clinical Practice Guideline index (PDF).
5. Prolonged Casualty Care Guidelines, 21 Dec 2021 (CPG ID 91, PDF).
6. Nursing Interventions, Wound Care, and Splint Management in Prolonged Casualty Care, 8 July 2025 (CPG ID 70 v1.1, PDF).
7. Baker K. "AI tools aim to help troops treat wounds." Business Insider, 7 Sept 2026.
8. Baker K. "Army special operations medics use AI to improve battlefield training." Business Insider, 24 July 2026.
9. Baker K. "The war in Ukraine is exposing the dangers of prolonged tourniquet use." Business Insider, July 2026.
10. "The Next War May Leave Wounded Troops Waiting Weeks for Evacuation." Business Insider, Aug 2026.
11. Kuziv R. LinkedIn post on TAAA/ALICE and SaMD regulatory work, 12 Aug 2026.
12. Neolens, "Prolonged Field Care" product page.
13. "AI assistant for the military: interview with the Neolens team." Militarnyi.
14. Wilson Intelligence Services, AEGIS MEDIC.
15. FDA, De Novo DEN250028 (SpectralMD DeepView System), granted 21 May 2026.
16. Spectral AI, "Spectral AI Receives FDA De Novo Clearance for DeepView® System for Burn Indication," GlobeNewswire, 26 May 2026.
17. FDA, 510(k) K233249 (APPRAISE-HRI), decision 5 April 2024.
18. USAMRDC, APPRAISE-HRI announcement, 11 April 2024.
19. FDA, De Novo DEN200022 (Fifth Eye, Analytic for Hemodynamic Instability), 1 March 2021.
20. Federal Register, classification of the Adjunctive Hemodynamic Indicator With Decision Point for Military Use, 27 Dec 2022.
21. FDA, Artificial Intelligence-Enabled Medical Devices list.
22. Congressional Research Service, IF13245, FDA regulation of AI-enabled medical devices, 10 June 2026. 22a. FDA, Considerations for the Regulation of Generative AI-Enabled Medical Devices, discussion paper, 18 August 2026, docket FDA-2026-N-7874, comments due 19 October 2026.
23. FDA, Marketing Submission Recommendations for a Predetermined Change Control Plan for AI-Enabled Device Software Functions, final guidance, August 2025.
24. Regulation (EU) 2026/1744 (Digital Omnibus on AI), OJ 24 July 2026 (PDF).
25. European Commission, AI Act implementation timeline.
26. Artificial Intelligence Strategy for the Department of War, January 2026 (PDF).
27. Brandi Vincent, "Feinberg orders major shakeup in Pentagon's AI enterprise," DefenseScoop, 15 August 2025.
28. Congressional Research Service, IN12615.
29. Exo, "Exo launches the first FDA-cleared ultrasound AI for detection of pleural effusion," June 2025.
30. Clarius Mobile Health 2025 results and FDA clearances.
31. "FDA Clears Clarius Ejection Fraction AI," 2 June 2026.
32. Connect Tech Sentry-X2 Mini, announced 12 May 2026.
33. Connect Tech Sentry-X2.
34. Aitech A230 Vortex.
35. 7StarLake NV200.
36. Willison S. "Trying out llama.cpp's new vision support."
37. "Field Testing of the Compensatory Reserve Measurement Monitoring Capability: Results From the 2024 Army Warfighter Expeditionary Experiment." Mil Med. 2025;190(Suppl 2):371–378.
38. "Validation of a Wearable Photoplethysmography-Based Sensor for Compensatory Reserve Measurement Monitoring in Simulated Human Hemorrhage." Sensors. 2026;26(8):2513.
39. FDA, De Novo DEN160020 (CipherOx CRI Tablet) (PDF).
40. Nature Medicine 2026, ChatGPT Health triage evaluation.
41. "Large language models versus emergency physicians in ED triage." J Clin Med. 2026;15(4):1512.
42. "Evaluation of 16 large language models on real-world emergency department triage." medRxiv, 28 May 2026.
43. Goddard K, Roudsari A, Wyatt JC. "Automation bias: a systematic review of frequency, effect mediators, and mitigators." JAMIA 2012;19(1):121–127.
44. Lyell D, Coiera E. "Automation bias and verification complexity: a systematic review." JAMIA 2017;24(2):423–431.
45. Budzyn K, et al. "Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy." Lancet Gastroenterol Hepatol 2025;10(10):896–903.
46. "Clinical validation of FDA-cleared AI/ML medical devices." PLOS Digital Health. 2026;5(8):e0001597.
47. "Artificial intelligence in triage of the injured in war conditions: a systematic review." J Sabzevar Univ Med Sci. 2026;33(3):1052–1068.
48. Mathais Q, Cungi P-J, Lamblin A, Dubourg O, Bordes J, Boussen S. "Artificial intelligence for battlefield triage in large-scale combat operations: Opportunities, limits, and ethical considerations." J Trauma Acute Care Surg. 2026;101:S177–S187.
49. "Artificial Intelligence in Military Medicine: Clinical, Operational, and Educational Applications." Rom J Mil Med. 2026;129(3):8.
50. National Academies of Sciences, Engineering, and Medicine. A National Trauma Care System (2016), Ch. 4.
51. "Saving Lives with Data." Military Review, Jan–Feb 2026.
52. DoD Trauma Registry Data Dictionary (external), v1.1 (PDF).
53. "Accuracy of injury coding in the Department of Defense Trauma Registry." Mil Med.
54. "An Analysis of the Shock Index and Pulse Pressure as a Predictor for Massive Transfusion and Death" (DoDTR, n=15,540).
55. "Shock Index as a Predictor of Massive Transfusion and Emergency Surgery on the Modern Battlefield" J Surg Res. 2020;256:112–118 (DoDTR 2008–2016, n=4,008).
56. "Shock index as predictor of massive transfusion and mortality in patients with trauma: a systematic review and meta-analysis" Crit Care. 2023;27:85 (n=670,728).
57. "An Assessment of Clinical Accuracy of Vital Sign-based Triage Tools Among U.S. and Coalition Forces." Mil Med. 2024;189(7–8):e1528–e1536.
58. DARPA Triage Challenge. 58a. DARPA, Triage Challenge events. 58b. DARPA, Triage Challenge finals data rules (PDF). 58c. DARPA, Triage Challenge data competition, Event 3.
59. "DART, MSAI triumph in DARPA Triage Challenge," 2025.
60. Microsoft, "Expanding healthcare AI models to new modalities and capabilities," HIMSS 2025.
61. Science and Technology Daily (PRC state media), 16 June 2026.
62. SAM Medical, SAM Splint product page.
63. Lundberg S, Lee S-I. "A Unified Approach to Interpreting Model Predictions." NeurIPS 2017.

Disclaimer: Views expressed are those of the author and do not reflect official positions of the U.S. Army, Department of Defense, or U.S. Government. Nothing on this page is medical advice, clinical guidance, or a substitute for training. The protocol summary in section 1 and the code in section 4 are descriptions of published doctrine written for software designers; neither is an instruction for treating a casualty, and the code has never been trained, validated, or evaluated for any use. Medical protocols must be applied only by qualified personnel acting within their scope of practice and under appropriate supervision and authority, and only against the current official guideline — Deployed Medicine and the Joint Trauma System — rather than this page.