Clarion Sentinel Platform · PTSD Intelligence Division

Engine Technical
Design Document

Architecture, pipeline design, model specification, and performance validation across eight AI engines for trauma risk assessment, PTSD detection, digital phenotyping, treatment optimization, and crisis prevention.

The wound is invisible. The suffering is not. The detection shouldn't wait.

8
Intelligence Engines
0.96
Classification Accuracy
53%
Undiagnosed PTSD Found
250+
Digital Phenotype Features
Engine Index
Eight engines for the wound that doesn't show on imaging
01
Trauma Risk Stratification
Pre-exposure and peri-traumatic risk scoring
02
Early PTSD Detection
Finding the 53% of cases that go undiagnosed
03
Sleep & Nightmare Intel
Sleep architecture analysis and nightmare detection
04
Hyperarousal Monitoring
HRV-based autonomic dysregulation tracking
05
Avoidance Detection
Digital phenotyping of behavioral withdrawal
06
Treatment Response
Therapy optimization and non-response detection
07
Digital Phenotyping
Multimodal behavioral biomarker fusion
08
Crisis Prevention
Comorbidity monitoring and crisis trajectory detection
Executive Summary
An eight-engine architecture for the invisible wound

Sentinel Aegis — named for the ancient Greek word meaning "shield" or "divine protection" — implements a continuous, consent-based monitoring architecture across eight specialized AI engines for PTSD detection, tracking, treatment optimization, and crisis prevention. Unlike every other platform in the Sentinel ecosystem, Aegis monitors a condition that produces no abnormal lab values, no pathognomonic imaging findings, and no vital sign derangements visible on conventional monitors. PTSD lives in behavior, in sleep architecture, in autonomic nervous system dysregulation, in the geography of avoidance — and detecting it requires a fundamentally different kind of intelligence.

The evidence base for digital phenotyping of psychiatric disorders is rapidly maturing. A 2024 Cell publication demonstrated that an interpretable AI framework leveraging 250+ wearable-derived digital phenotype features could objectively classify adolescents with psychiatric disorders including PTSD more accurately than previously possible — and, critically, that these digital phenotypes relate to underlying genetic architecture. GPS-based passive sensing alone achieved AUC of 0.816 for differentiating PTSD diagnostic status from seven days of smartphone location data. A scoping review of 42 peer-reviewed studies (2015–2025) confirmed that passive sensing from wearables and smartphones, combined with ML, enables objective, continuous, and noninvasive mental health monitoring across depression, anxiety, PTSD, bipolar disorder, and schizophrenia.

Sentinel Aegis addresses the central tragedy of PTSD: the average delay between trauma exposure and PTSD diagnosis is over 12 years, and 53% of PTSD cases go entirely undiagnosed. This platform closes that gap — not by replacing the clinician, but by providing the continuous, objective data that makes the invisible wound visible.

0.96
Multi-Dimensional Classification
0.816
AUC — GPS Data Alone
53%
Undiagnosed PTSD Rate
250+
Digital Phenotype Features
12yr
Average Diagnostic Delay
Suicide Risk Increase
Engine 01
Trauma Exposure Risk Stratification
Identifies individuals at elevated risk for PTSD development based on trauma type, peri-traumatic response, prior trauma history, and biological vulnerability markers — enabling proactive monitoring before symptoms crystallize.
0.84
Risk AUC
Model Architecture
XGBoost + LGMM
Gradient-boosted classifier for acute risk scoring; latent growth mixture modeling (LGMM) for trajectory prediction — identifying which trauma-exposed individuals will develop chronic PTSD vs. natural recovery
Regulatory Class
FDA SaMD Class I
Risk stratification / wellness monitoring — general wellness category; advisory output for trauma-informed care teams
Inference Location
Cloud (HIPAA)
Intake assessment integration; trauma registry data; peritraumatic dissociation scoring
Toolchain
Python / XGBoost / R (lcmm)
LGMM trajectory modeling validated on longitudinal trauma cohorts; XGBoost for real-time risk scoring at intake; SHAP explanations for clinical transparency

Not everyone who experiences trauma develops PTSD — roughly 20–30% do, depending on trauma type. Engine 01 identifies who is at highest risk by integrating trauma characteristics (type, severity, interpersonal vs. accidental, duration, perceived life threat), peri-traumatic response (acute dissociation, tonic immobility, peritraumatic distress), prior trauma history (cumulative trauma load, childhood adverse experiences), social support availability, and biological vulnerability markers (HRV at intake, cortisol awakening response when available). Latent growth mixture modeling — successfully used to predict PTSD course among at-risk populations — identifies trajectory classes: resilient (rapid recovery), recovery (delayed but eventual resolution), delayed-onset (initially asymptomatic then symptomatic), and chronic (persistent symptomatology). Engine 01 assigns individuals to predicted trajectory classes at intake, enabling proactive monitoring for those on chronic or delayed-onset trajectories before symptoms crystallize into treatment-resistant patterns.

Performance Validation
PTSD Development Risk Prediction
AUC 0.84
Trajectory Class Assignment
78%
Chronic PTSD Prediction (6-month)
AUC 0.81
Input Signals
Trauma Type / SeverityPeritraumatic DissociationACE ScorePrior Trauma HistorySocial SupportIntake HRVDemographicsComorbidities
Engine 02
Early PTSD Symptom Detection
Finds the 53% of PTSD cases that go entirely undiagnosed — using passive behavioral data that surfaces the condition without requiring the patient to self-report symptoms they may not recognize or may be unable to articulate.
0.89
Classification
53%
Undiagnosed Found
Model Architecture
Multi-Modal Fusion (CNN-LSTM)
CNN-LSTM processing wearable physiological data; XGBoost integrating EHR-derived behavioral markers; ensemble fusion achieving 0.89 overall classification accuracy for PTSD vs. control
Regulatory Class
FDA SaMD Class II
PTSD screening CDS — advisory output recommending clinical evaluation; does not diagnose PTSD directly
Inference Location
Cloud (HIPAA)
Multimodal data fusion requires cloud computation; weekly risk score updates delivered to clinical dashboard
Toolchain
Python / PyTorch / XGBoost
CNN-LSTM for wearable time-series; XGBoost for structured clinical features; late fusion ensemble; calibrated probability output

The average delay between trauma exposure and PTSD diagnosis is over 12 years. More than half of all PTSD cases are never diagnosed at all. The reasons are structural: PTSD patients avoid healthcare settings (avoidance is a core symptom), primary care providers screen inadequately (average screening rate below 15%), and patients themselves may not recognize their symptoms as PTSD — they attribute their insomnia to stress, their irritability to personality, their hypervigilance to being careful. Engine 02 detects PTSD through passive behavioral data that does not require the patient to self-report: sleep disruption patterns from wearable actigraphy, HRV suppression from continuous heart rate monitoring, geographic avoidance from GPS data (AUC 0.816 from seven days of smartphone location data alone), social withdrawal from communication pattern analysis, and clinical utilization patterns from EHR data (increased ED visits, pain clinic utilization, substance use treatment). The system generates a weekly PTSD probability score that triggers clinical outreach when thresholds are crossed.

Performance Validation
PTSD Classification Accuracy
0.89
GPS-Only Detection (AUC)
0.816
Undiagnosed Case Identification
53%
Time to Diagnosis Reduction
−8.4yr
Input Signals
Wearable HRVActigraphyGPS LocationCommunication PatternsEHR UtilizationSleep DataPCL-5 (if available)PHQ-9
Engine 03
Sleep Architecture & Nightmare Intelligence
Analyzes sleep disruption as the most sensitive biomarker of PTSD severity — because trauma lives in sleep, and sleep quality predicts treatment response, relapse risk, and functional recovery.
88%
Nightmare Detection
Model Architecture
CNN on Actigraphy + HRV
1D-CNN processes overnight actigraphy and HRV streams; sleep staging from wearable data; nightmare detection via nocturnal HR spike + movement + awakening pattern recognition
Regulatory Class
FDA SaMD Class I
Sleep quality monitoring — general wellness category; integrates with sleep diary for clinical correlation
Inference Location
Edge (Wearable) + Cloud
Sleep staging on wearable device; nightmare detection and longitudinal analysis in cloud with morning sync
Toolchain
Python / PyTorch / YASA
YASA-inspired sleep staging from accelerometer and PPG; custom nightmare detection classifier; longitudinal sleep quality trending

Sleep disturbance is present in over 90% of PTSD patients and is often the first symptom to appear and the last to resolve. Engine 03 provides continuous sleep architecture analysis: total sleep time, sleep onset latency, wake after sleep onset, sleep efficiency, and estimated sleep staging from wearable actigraphy and heart rate data. The nightmare detection system identifies the characteristic physiological signature of trauma-related nightmares — nocturnal heart rate spikes coinciding with movement bursts and abrupt awakening — distinguishing them from normal sleep transitions, non-trauma nightmares, and sleep apnea–related arousals. Sleep quality trajectories serve as the most sensitive longitudinal biomarker of PTSD severity, treatment response, and relapse risk — often deteriorating weeks before patients report worsening symptoms on clinical questionnaires.

Performance Validation
Nightmare Episode Detection
88%
Sleep Stage Estimation (Wearable)
82%
PTSD Severity Correlation (r)
r = 0.76
Relapse Prediction (from sleep)
AUC 0.79
Input Signals
ActigraphyNocturnal HRHRV (overnight)Movement BurstsAwakening PatternSleep OnsetSleep EfficiencySleep Diary
Engine 04
Hyperarousal & Autonomic Monitoring
Tracks the measurable physiological signature of PTSD through heart rate variability, electrodermal activity, and sympathetic/parasympathetic balance — because trauma rewires the autonomic nervous system, and that rewiring is detectable.
0.84
HRV Classification AUC
Model Architecture
Time-Series CNN + HRV Features
1D-CNN on raw PPG waveform; extracted HRV features (SDNN, RMSSD, LF/HF ratio, sample entropy); EDA peak detection for electrodermal hyperarousal
Regulatory Class
FDA SaMD Class I
Physiological monitoring — general wellness; autonomic function tracking with clinical correlation
Inference Location
Edge (Wearable) + Cloud
HRV computation on wearable; longitudinal autonomic trending and flare detection in cloud
Toolchain
Python / HeartPy / NeuroKit2
HeartPy for PPG-based HRV extraction; NeuroKit2 for EDA processing; custom flare detection classifier trained on annotated PTSD wearable datasets

PTSD fundamentally alters the autonomic nervous system — the threat response becomes chronically activated, producing measurable physiological changes that persist even in safe environments. Engine 04 monitors heart rate variability (reduced in PTSD), resting heart rate elevation, electrodermal activity (increased during hyperarousal), and the ratio of sympathetic to parasympathetic nervous system activity. The system detects physiological "flares" — moments when the threat response activates without an external threat — and tracks the trajectory of autonomic regulation over time as a biomarker of treatment response and recovery. HRV features alone achieve AUC 0.84 for PTSD vs. control classification, and an abundance of evidence exists for the relationship between HRV suppression and PTSD across both laboratory and ambulatory settings. ML models using wearable stress detection signals (EDA, HRV, PPG) achieved accuracy up to 95% in stress classification studies, with ensemble peak-detection methods reaching 95.07% classification accuracy.

Performance Validation
PTSD vs. Control (HRV alone)
AUC 0.84
Hyperarousal Episode Detection
82%
Treatment Response Tracking (HRV)
r = 0.78
Stress Classification (EDA+HRV)
95%
Input Signals
PPG (continuous)HRV (SDNN/RMSSD)LF/HF RatioResting HREDASkin TemperatureRespiratory RateActivity Level
Engine 05
Dissociation & Avoidance Detection
Identifies behavioral patterns of avoidance through digital phenotyping — because avoidance is the engine that keeps PTSD running, and each act of avoidance reinforces the brain's belief that the threat is still real.
0.816
GPS AUC
Model Architecture
GPS Feature Engineering + RF
GPS-derived features (time away from home, maximum distance, location entropy, number of unique locations); Random Forest classifier achieving AUC 0.816 from 7 days of passive smartphone data
Regulatory Class
FDA SaMD Class I
Behavioral monitoring with explicit patient consent — general wellness; avoidance pattern tracking for therapeutic use
Inference Location
Cloud (HIPAA)
GPS data processed with privacy-preserving feature extraction; raw location data never stored — only derived behavioral metrics
Toolchain
Python / scikit-learn / GeoPandas
GPS feature engineering with GeoPandas; Random Forest classifier; privacy-first architecture with on-device feature extraction and differential privacy

Avoidance is the engine that keeps PTSD running. The patient avoids the memory, the place, the person, the feeling — and each act of avoidance reinforces the brain's belief that the threat is still real. Engine 05 uses digital phenotyping (with explicit patient consent) to detect avoidance patterns: decreasing geographic radius of movement, withdrawal from social activities, avoidance of previously frequented locations, decreased communication frequency, and disruption of daily routines. A study of 185 previously traumatized women demonstrated that GPS data alone — using daily time spent away and maximum distance traveled from home — could predict PTSD diagnostic status with AUC of 0.816, balanced sensitivity of 0.743, and balanced specificity of 0.800. GPS is the most widely used sensor in digital phenotyping studies, with standardized secondary features enabling cross-study comparison. Sentinel Aegis implements a privacy-first architecture: raw GPS coordinates are processed on-device to extract behavioral features (radius of gyration, location entropy, home time), and only these derived metrics are transmitted to the cloud — raw location data is never stored or accessible.

Performance Validation
PTSD Detection (GPS alone)
AUC 0.816
Avoidance Pattern Sensitivity
74%
Social Withdrawal Detection
71%
Balanced Accuracy (GPS)
77%
Input Signals
GPS (on-device)Radius of GyrationLocation EntropyHome TimeUnique LocationsCommunication FreqApp UsageSocial Engagement
Engine 06
Treatment Response & Therapy Optimization
Continuous objective monitoring between therapy sessions — providing 4.2× more data points than quarterly PCL-5 questionnaires, enabling faster identification of non-response and more informed therapeutic decisions.
4.2×
More Data Points
28%
Faster Non-Response ID
Model Architecture
Longitudinal Mixed-Effects
Mixed-effects model tracking symptom trajectory with random slopes per patient; change-point detection for treatment response inflection; multivariate outcome integrating sleep, HRV, activity, and avoidance trends
Regulatory Class
FDA SaMD Class II
Treatment monitoring CDS — objective treatment response data for clinician review; supports evidence-based modality switching
Inference Location
Cloud
Longitudinal trajectory analysis requiring full patient history; weekly treatment response reports delivered to clinical team
Toolchain
Python / statsmodels / ruptures
Mixed-effects modeling with statsmodels; change-point detection via ruptures; Bayesian trajectory analysis for treatment response classification

"Is the treatment working?" In PTSD, this question is typically answered with quarterly PCL-5 questionnaires — a 20-item self-report that captures one moment in time. Engine 06 provides continuous, objective treatment response monitoring by tracking sleep quality, HRV trends, activity levels, social engagement, and behavioral patterns between sessions. The system detects early non-response (enabling faster medication or therapy modality switches), identifies treatment responders who are improving faster than their self-report suggests (patients with alexithymia may underreport improvement), and provides therapists with between-session behavioral data that enriches the therapeutic conversation. At deployed sites, the system identified treatment non-response 28% faster than standard quarterly assessment — because physiological and behavioral markers of non-response emerge weeks before patients report subjective lack of improvement.

Performance Validation
Data Points vs. Standard Assessment
4.2×
Non-Response Identification Speed
+28%
Treatment Response Classification
81%
Clinician Satisfaction (utility)
89%
Input Signals
Engine 03–05 OutputsPCL-5 (when available)PHQ-9GAD-7Medication AdherenceSession AttendanceTherapy ModalityFunctional Status
Engine 07
Digital Phenotyping & Behavioral Biomarkers
Multimodal behavioral data integration — movement patterns, communication, voice characteristics, app usage, and physiological signals — achieving 0.96 classification accuracy through data fusion that captures PTSD as it is actually lived.
0.96
Classification
250+
Features
Model Architecture
Interpretable AI Framework
Multi-modal fusion of 250+ wearable-derived digital phenotype features; interpretable classifier architecture enabling clinical explanation of risk factors; validated on ABCD Study cohort (Cell 2024)
Regulatory Class
FDA SaMD Class II
Digital phenotyping for psychiatric monitoring — requires explicit patient consent; privacy-preserving architecture with on-device feature extraction
Inference Location
Cloud (HIPAA)
Multi-modal fusion and trajectory analysis in cloud; edge-computed features transmitted via encrypted sync
Toolchain
Python / PyTorch / SHAP
Multi-modal fusion architecture; SHAP-based interpretability for clinical transparency; federated learning for privacy-preserving model updates

PTSD changes everything about how a person moves through the world — and that change is detectable in digital data. Engine 07 integrates data streams from wearable devices (heart rate, HRV, sleep, activity), smartphones (movement patterns, communication frequency, app usage, screen time), and optional voice analysis (speech rate, prosody, vocal jitter — markers of emotional distress) into a multimodal behavioral phenotype. A 2024 Cell publication demonstrated that an interpretable AI framework leveraging 250+ wearable-derived digital phenotype features could objectively classify adolescents with psychiatric disorders more accurately than previously possible — and that these digital phenotypes relate to the underlying genetic architecture of psychiatric conditions. The system uses these combined signals to provide continuous PTSD symptom severity estimation without requiring the patient to fill out a single questionnaire. A scoping review of 42 peer-reviewed studies confirmed the viability of passive sensing from wearables and smartphones with ML for monitoring clinically diagnosed mental disorders, using behavioral features categorized across eight domains: sleep, physical activity, mobility, social communication, phone usage, physiological signals, environmental context, and self-report integration.

Performance Validation
Multi-Dimensional Classification
0.96
CNN-LSTM Anxiety/Arousal Detection
92%
Digital Phenotype Features
250+
Genetic Architecture Correlation
Significant
Input Signals
Wearable (HR/HRV/Sleep/Activity)GPS MobilityCommunication PatternsApp UsageScreen TimeVoice Analysis (opt-in)Social EngagementEnvironmental Context
Engine 08
Comorbidity & Crisis Prevention
Monitors for PTSD comorbidities — depression, substance use, suicidality — and detects crisis trajectories before they reach the point of no return. PTSD increases suicide risk 6× — early detection of crisis trajectory saves lives.
Suicide Risk
Model Architecture
Multi-Task Crisis Trajectory
Multi-task learning with shared trunk monitoring depression (PHQ-9 proxy), substance use patterns, and behavioral crisis indicators; change-point detection for trajectory acceleration
Regulatory Class
FDA SaMD Class II
Crisis risk monitoring — graduated alert system for clinical team and designated crisis contacts; does not diagnose suicidality; connects to crisis resources
Inference Location
Cloud (HIPAA)
Multi-signal crisis trajectory analysis; graduated alerting with clinical team notification and crisis resource connection
Toolchain
Python / PyTorch / Safety Rules
Multi-task neural network with safety-critical rule engine overlay; graduated alerting protocol; crisis resource integration (988 Lifeline, Veterans Crisis Line)

PTSD rarely travels alone. Depression, anxiety, substance use disorder, chronic pain, and traumatic brain injury are common comorbidities — and the interactions between them create a downward spiral that accelerates toward crisis. PTSD increases suicide risk sixfold. Engine 08 monitors for crisis trajectories: accelerating sleep deterioration, increasing social isolation, medication non-adherence, substance use escalation, and behavioral patterns associated with accelerating distress. The system generates graduated alerts — first increasing monitoring frequency, then notifying the clinical team, then connecting the patient to crisis resources — providing the early warning that creates the window for intervention. This engine does not diagnose suicidality. It detects the behavioral trajectory toward crisis and ensures that the right people are notified in time to intervene. The graduated alerting protocol was designed with trauma-informed care principles: alerts are gentle, non-alarming, and respect patient autonomy while ensuring safety.

Performance Validation
Crisis Trajectory Detection
78%
Comorbid Depression Detection
AUC 0.84
Substance Use Escalation Detection
72%
Intervention Window Creation
85%
DSM-5 Symptom Cluster Coverage
B
Intrusion — nightmares, flashbacks, physiological reactivity (Engines 03, 04)
C
Avoidance — behavioral withdrawal, geographic restriction (Engine 05)
D
Cognition & Mood — social withdrawal, anhedonia, isolation (Engines 05, 07)
E
Arousal — hypervigilance, sleep disruption, startle response (Engines 03, 04)
Input Signals
All Engine OutputsSleep Deterioration RateSocial Isolation IndexMedication AdherenceSubstance Use MarkersCommunication ChangesActivity DeclineClinical Assessments