Clarion Sentinel Platform · Medication Safety Division

Engine Technical
Design Document

Architecture, pipeline design, model specification, and performance validation across eight AI engines for drug-drug interaction prediction, adverse event detection, polypharmacy risk stratification, organ-function dose adjustment, anticoagulant safety, opioid surveillance, prescribing cascade prevention, and medication reconciliation.

Engines
8 Medication Safety Systems
DDI Classification
Knowledge Graph · GNN · AUC 0.92
Alert Fatigue
78% Reduction · 99.2% Sensitivity
Classification
Confidential — Internal Use Only
Contents
Eight Engines
01
Drug-Drug Interaction Prediction
Knowledge graph + GNN DDI classification with context-aware severity stratification eliminating 78% of alert fatigue
02
Adverse Drug Event Early Detection
Multi-modal ADE prediction using EHR time-series, lab trajectories, and NLP clinical notes — 12–36 hour lead time
03
Polypharmacy Risk Stratification
Anticholinergic burden, Drug Burden Index, fall risk, and AI-guided deprescribing — AUC 0.87
04
Renal & Hepatic Dose Adjustment
Real-time GFR/CrCl/Child-Pugh monitoring with automatic dose recalculation and nephrotoxin detection
05
Anticoagulant Safety Intelligence
INR trajectory prediction, DOAC interaction surveillance, HAS-BLED monitoring, and bridging guidance
06
Opioid Risk & Accumulation Detection
MME calculation, accumulation trajectory, respiratory depression risk, and PDMP integration
07
Prescribing Cascade Prevention
Pattern recognition across 340 known cascades — detecting when new prescriptions treat side effects of existing ones
08
Medication Reconciliation Intelligence
AI-powered reconciliation at every care transition — admission, transfer, discharge — with discrepancy detection
Executive Summary
System Architecture Overview
Sentinel Pharma deploys eight interconnected AI engines that create a continuous medication safety surveillance layer across every hospitalized patient. The platform addresses the fourth leading cause of death in hospitalized patients — adverse drug events harm more than 1.3 million people annually in the US alone, costing $42 billion in preventable healthcare expenditure. The fundamental architectural innovation is context-aware alert stratification: current clinical decision support systems generate so many low-severity drug-drug interaction alerts that clinicians override approximately 90% of them, including the critical ones. Sentinel Pharma uses a knowledge graph architecture with graph neural network classification (AUC 0.92 for ADR causality assessment) to stratify interaction severity based on each patient's specific clinical context — renal function, hepatic status, age, weight, comorbidities, and concurrent medications.
The architecture integrates three AI paradigms: knowledge graphs that encode pharmacological relationships between drugs, adverse events, CYP450 enzymes, and patient characteristics as nodes and edges, enabling relational reasoning that flat databases cannot perform; Transformer-based NLP models (DrugBERT, BioBERT) that extract emerging interaction signals from clinical documentation and medical literature before they appear in structured databases; and time-series ML models that monitor lab value trajectories (creatinine rising after NSAID initiation, platelets falling on heparin, INR widening after antibiotic addition to warfarin) to detect adverse drug events 12–36 hours before clinical manifestation. Multi-modal approaches combining EHR, literature, and adverse event reporting data provide lead-time improvements of 7–22 months for detecting ADRs relative to regulatory labeling revision dates.
78%
Alert Fatigue Reduction
99.2%
Sensitivity for Critical DDIs
0.92
AUC for ADR Classification
12-36hr
ADE Early Detection Lead Time
Engine 01
Drug-Drug Interaction Prediction
90% of DDI alerts are overridden — because alert fatigue kills as many patients as missed interactions

Current DDI alerting systems generate so many low-severity alerts that clinicians override approximately 90% of them — including the critical ones that could save lives. The fundamental problem is not detection sensitivity (existing systems detect most DDIs) but severity stratification (existing systems cannot distinguish a theoretical interaction between two drugs that a patient tolerates well from a life-threatening interaction in a patient with compromised organ function). Sentinel Pharma uses AI to stratify interaction severity based on each patient's specific clinical context: renal function, hepatic status, CYP450 genotype (where available), age, weight, comorbidities, and the complete concurrent medication list. The system suppresses theoretical interactions while escalating truly dangerous combinations — reducing alert fatigue 78% while maintaining 99.2% sensitivity for clinically significant DDIs and increasing clinician alert acceptance rates 4.2× from baseline.

The DDI prediction architecture uses a pharmacological knowledge graph with 340,000+ drug-drug, drug-enzyme, drug-transporter, and drug-condition relationship edges, classified by a graph neural network that achieves AUC of 0.97 for DDI prediction and AUC of 0.92 for adverse drug reaction causality assessment. Transformer-based models (DrugBERT) extract emerging interaction signals from clinical literature and FDA adverse event reports, enabling detection of novel DDIs before they appear in structured interaction databases — providing a lead-time advantage of 7–22 months over traditional pharmacovigilance methods.

78%
Reduction in low-priority DDI alerts (alert fatigue elimination)
99.2%
Sensitivity for clinically significant drug-drug interactions
4.2×
Increase in clinician alert acceptance rate
0.97
AUC for GNN-based DDI prediction
7-22mo
Lead time for novel DDI detection vs. regulatory labeling
DDI Detection Pipeline
STAGE 01
Medication Ingestion
Real-time integration with EHR, pharmacy systems, CPOE, and medication administration records. Complete medication list including OTC, supplements, and home medications.
EHRCPOEMAR
STAGE 02
Knowledge Graph Query
Each medication pair queried against 340K+ relationship edges: CYP450 substrate/inhibitor/inducer relationships, QT prolongation stacking, serotonergic risk, bleeding risk summation.
KGCYP450QTc
STAGE 03
Patient Context Scoring
GNN classifies interaction severity adjusted for patient-specific factors: GFR, hepatic function, age, body composition, genomics (CYP2D6/2C19 when available), and concurrent drugs.
GNNPatient Context
STAGE 04
Alert Stratification
Three-tier classification: suppress (theoretical/well-tolerated), inform (moderate with monitoring recommendation), alert (critical with immediate action required). Only critical alerts interrupt workflow.
Stratify3-Tier
STAGE 05
Clinical Decision Support
Critical alerts delivered with: specific mechanism of harm, patient-specific risk factors, recommended alternative medications, monitoring parameters, and one-click acceptance/override with reason capture.
CDSSMART on FHIR
Knowledge Graph Architecture

The pharmacological knowledge graph encodes 340,000+ relationships across five entity types: drugs (24,000+ compounds), enzymes (CYP450 family, UGT, transporter proteins), adverse events (8,400+ MedDRA preferred terms), patient conditions (ICD-10 comorbidities), and pharmacogenomic variants (CYP2D6, CYP2C19, CYP3A4, VKORC1 alleles). Edges encode relationship types: substrate-of, inhibitor-of, inducer-of, increases-risk-of, contraindicated-with, and dose-adjustment-required-for. The graph neural network learns to predict interaction severity by propagating information through the graph — a drug that inhibits CYP3A4 interacts with every substrate of CYP3A4, but the clinical significance depends on the substrate's therapeutic index, the patient's alternative clearance pathways, and the availability of monitoring parameters to detect early toxicity. Knowledge graph-based methods have achieved AUC of 0.92 in classifying known causes of ADRs, substantially outperforming traditional statistical methods that typically achieve AUCs of 0.70–0.80 for similar tasks.

DrugBERT Signal Detection

The novel DDI detection module uses DrugBERT — a Transformer model fine-tuned on biomedical literature — to extract emerging interaction signals from three data streams: (1) FDA Adverse Event Reporting System (FAERS) spontaneous reports, where disproportionality analysis identifies drug pairs with unexpectedly high co-reported adverse events; (2) medical literature (MEDLINE, 38 million citations), where NLP pipelines extract drug-drug interaction mentions from case reports, pharmacokinetic studies, and systematic reviews; (3) clinical documentation (de-identified clinical notes from deployed sites), where DrugBERT identifies implicit DDI references in physician documentation — "patient developed rash after adding [Drug B] to [Drug A]" — that represent real-world interaction signals preceding formal pharmacovigilance recognition. Multi-modal approaches combining these sources provide AUC improvements of 0.04–0.09 compared to single-source methods, with a lead-time of 7–22 months for detecting ADRs relative to labeling revision dates.

Clinical Impact at Deployed Sites
78%
Fewer low-priority alerts
4.2×
Alert acceptance rate improvement
34%
ADE rate reduction
23
Life-threatening DDIs caught in year one
Engine 02
Adverse Drug Event Early Detection
A rising creatinine. A dropping platelet count. A widening INR. The signals are there — 12 to 36 hours early

Most adverse drug events announce themselves before they become emergencies — through subtle lab value changes, vital sign trends, and clinical documentation patterns that individually appear unremarkable but collectively signal an emerging drug-related injury. A rising creatinine after starting an NSAID. A dropping platelet count on heparin. A widening INR after adding an antibiotic to warfarin. Engine 02 monitors these signals continuously and generates proactive alerts when trajectories suggest an emerging ADE — giving clinicians the 12–36 hour intervention window before the patient codes, bleeds, or seizes. A meta-analysis of 59 studies covering 15 drugs and 15 ADEs found that ML models achieved an average AUC of 0.77 for ADE prediction from EHR data, with individual disease-specific models reaching AUC of 0.89–0.90 for hepatotoxicity prediction and 0.94 for ADR classification using drug-target interaction data.

12-36hr
Earlier ADE detection vs. standard clinical recognition
72%
Of ADEs detected before clinical manifestation
0.87
AUC for ADE risk prediction across medication classes
Time-Series ADE Detection

The ADE detection model uses a multi-channel LSTM that processes parallel time series for each patient: laboratory values (creatinine, platelets, INR, liver enzymes, electrolytes), vital signs (heart rate, blood pressure, respiratory rate, temperature), and medication administration events. The model learns the temporal signatures of drug-induced organ injury — the characteristic creatinine trajectory that follows NSAID nephrotoxicity (gradual rise over 3–5 days) differs from the trajectory of contrast-induced nephropathy (sharp rise within 24–48 hours) and from the trajectory of aminoglycoside toxicity (dose-dependent cumulative rise). By recognizing these drug-specific injury patterns, the system achieves 72% sensitivity for detecting ADEs before they manifest clinically, with an average lead time of 12–36 hours — the critical window where dose adjustment, drug discontinuation, or protective intervention can prevent progression to organ failure.

NLP Clinical Documentation Mining

The clinical documentation module uses BioBERT fine-tuned on 2.8 million de-identified clinical notes to extract implicit ADE signals from physician and nursing documentation. The model identifies mentions of symptoms that may represent unrecognized adverse drug effects: "patient reports new-onset dizziness" following statin initiation, "increased bruising noted" on dual antiplatelet therapy, "confusion and agitation" in an elderly patient with recent anticholinergic medication addition. These NLP-extracted signals are correlated with the patient's medication timeline to identify temporal associations that suggest drug causality. The system applies the Naranjo ADR probability scale algorithmically — scoring temporality, dechallenge, rechallenge, alternative causes, and dose-response relationships — to generate causality probability scores that prioritize investigation for the most likely drug-event relationships.

Engine 03
Polypharmacy Risk Stratification
Not the number of medications — the cumulative physiological burden they impose

Polypharmacy is not merely the count of medications — it is the cumulative physiological burden they impose. Engine 03 calculates anticholinergic burden scores, sedation risk indices, fall risk contributions, cognitive impairment risk, and Drug Burden Index for every patient in real time. The system identifies patients at highest risk for polypharmacy-related adverse events and generates prioritized deprescribing recommendations — medications that can be safely discontinued, doses that can be reduced, and therapeutic duplications that should be resolved. ML models achieve AUC exceeding 0.87 for polypharmacy adverse event prediction, and the STOPP/START criteria integration has proven effective in minimizing potentially inappropriate medications in geriatric populations. A geriatric care network deployment achieved an average of 2.4 medications safely deprescribed per high-risk patient, reducing polypharmacy-related hospitalizations 34% and falls 28%.

0.87
AUC for polypharmacy adverse event prediction
2.4
Average medications safely deprescribed per high-risk patient
34%
Reduction in polypharmacy-related hospitalizations
28%
Reduction in falls after anticholinergic burden reduction
Burden Scoring Architecture

The polypharmacy engine computes five composite burden indices simultaneously: (1) Anticholinergic Cognitive Burden Scale — summing anticholinergic potency scores across all medications, with scores ≥3 associated with significant cognitive impairment risk; (2) Sedation Risk Index — cumulative sedating effect from opioids, benzodiazepines, antihistamines, antipsychotics, and muscle relaxants; (3) Drug Burden Index (DBI) — composite score incorporating both anticholinergic and sedative burden, validated as an independent predictor of falls, hip fractures, and functional decline; (4) QTc Prolongation Risk — cumulative effect of QT-prolonging medications adjusted for electrolyte status, heart rate, and baseline QTc; (5) Serotonergic Load — risk assessment for serotonin syndrome from combinations of SSRIs, SNRIs, triptans, tramadol, linezolid, and other serotonergic agents. Each index operates continuously as a running total that updates immediately when any medication is added, removed, or dose-adjusted.

Deprescribing Intelligence

The deprescribing module generates prioritized recommendations for medication reduction using a clinical evidence graph that encodes which medications can be safely discontinued, what tapering protocols are required (abrupt discontinuation vs. gradual reduction), what monitoring is needed during discontinuation, and what the expected timeline for benefit realization is. The system integrates the STOPP/START criteria (Screening Tool of Older People's Prescriptions / Screening Tool to Alert to Right Treatment) to identify both potentially inappropriate medications that should be stopped and evidence-based medications that are missing. Recommendations are ranked by expected harm-reduction value: medications contributing most to the patient's aggregate burden scores are prioritized for deprescribing review, enabling clinicians to focus their limited medication review time on the changes most likely to improve patient outcomes.

Engine 04–05
Renal & Hepatic Dose Adjustment · Anticoagulant Safety
25% of AKI-related ADEs involve unadjusted doses · Anticoagulants cause 33% of emergency ADE hospitalizations

Engine 04 continuously monitors GFR, creatinine clearance, and hepatic function markers (Child-Pugh score, MELD), cross-referencing each patient's medication list against organ-specific dosing guidelines. When organ function declines below dosing thresholds, the system generates immediate dose adjustment recommendations — specifying the exact dose reduction, interval extension, or drug substitution required. The system also detects nephrotoxic and hepatotoxic medication combinations that accelerate organ damage. Engine 05 provides continuous anticoagulant safety monitoring: predicting INR trajectories from dietary changes, new medications, and illness patterns; detecting DOAC interactions that increase bleeding risk; monitoring HAS-BLED scores dynamically; alerting to dual antiplatelet/anticoagulant therapy without clear indication; and guiding periprocedural bridging decisions. Anticoagulants are responsible for one-third of emergency hospitalizations for adverse drug events — the single most dangerous class of medications in common clinical use.

94%
Of renal-dose adjustments identified before ADE occurrence
42%
Reduction in renal-dosing-related adverse events
33%
Of emergency ADE hospitalizations caused by anticoagulants
44%
Reduction in anticoagulant bleeding events at deployed sites
Organ-Function Dose Engine

The renal dose adjustment engine monitors GFR (calculated via CKD-EPI equation using serum creatinine and cystatin C when available) and creatinine clearance (Cockcroft-Gault for drugs with CrCl-based dosing recommendations). When GFR declines below drug-specific thresholds, the system immediately cross-references the patient's medication list against a curated database of 2,400+ renally-eliminated medications with dose adjustment recommendations from FDA prescribing information, clinical pharmacology references, and renal dosing guidelines. The system handles the complexity that many medications have different adjustment thresholds — metformin contraindicated below GFR 30, gabapentin requires adjustment below GFR 60, and vancomycin requires individualized dosing based on actual clearance rather than threshold-based adjustment. Hepatic dose adjustments follow a parallel architecture using Child-Pugh classification and MELD score.

Anticoagulant Safety Architecture

The anticoagulant safety engine maintains a dedicated monitoring pathway for every anticoagulated patient: warfarin patients receive INR trajectory prediction using an LSTM model trained on 180,000 INR measurements that predicts the 48-hour INR trajectory from current INR, recent dietary patterns (vitamin K intake), new medications (CYP2C9 inhibitors/inducers), and illness patterns (febrile illness accelerates warfarin metabolism). DOAC patients (apixaban, rivarelbaan, edoxaban, dabigatran) receive interaction surveillance focused on P-glycoprotein and CYP3A4 modulators that alter DOAC levels, renal function monitoring (DOACs have varying degrees of renal elimination), and bleeding risk assessment using dynamically updated HAS-BLED scores. A health system deployment across 14,000 anticoagulated patients reduced anticoagulant-related bleeding events by 44% — with the most dramatic impact in patients on concurrent anticoagulant-antiplatelet therapy.

Engine 06–08
Opioid Safety · Prescribing Cascades · Medication Reconciliation
MME accumulation · 340 cascade patterns · Every care transition audited

Engine 06 calculates total morphine milligram equivalents (MME) across all opioid prescriptions, monitors for accumulation trajectories in patients with declining renal or hepatic function, assesses respiratory depression risk using a composite score (opioid dose, concurrent sedatives, sleep apnea, COPD, age), and integrates with state Prescription Drug Monitoring Programs (PDMPs) to identify patients receiving opioids from multiple prescribers. Engine 07 detects prescribing cascades — the clinical pattern where a new drug is prescribed to treat the side effect of an existing drug, which in turn causes its own side effects, triggering yet another prescription. The system monitors 340 known cascade patterns, identifying when a new prescription likely represents a cascade rather than a new condition. Engine 08 provides AI-powered medication reconciliation at every care transition — admission, unit transfer, and discharge — using NLP to reconcile discrepancies between home medication lists, inpatient orders, and discharge prescriptions.

MME
Real-time morphine milligram equivalent tracking across all opioid sources
340
Known prescribing cascade patterns monitored continuously
15%
Of new prescriptions in elderly patients are cascade prescriptions
Every
Care transition reconciled — admission, transfer, discharge
Opioid Accumulation Model

The opioid safety engine maintains a pharmacokinetic model for each patient that tracks not just current MME dose but the predicted plasma opioid level trajectory based on drug-specific half-lives, metabolic pathways, and organ function. Methadone, for example, has a half-life of 8–59 hours (highly variable and dose-dependent) — meaning that steady-state plasma levels may not be reached for 5–7 days after dose changes. The system predicts accumulation risk by modeling the approach to steady state, alerting clinicians when predicted peak plasma concentrations approach respiratory depression thresholds. Concurrent administration of benzodiazepines, gabapentinoids, or muscle relaxants multiplicatively increases respiratory depression risk — the system computes composite sedation scores that account for the pharmacodynamic interaction between opioids and co-sedatives, which traditional MME-only monitoring fails to capture.

Cascade Detection Architecture

The prescribing cascade detection engine uses a temporal pattern matching algorithm that evaluates every new prescription against the patient's medication history and symptom timeline. The system maintains a curated database of 340 known cascade patterns — for example: (1) NSAID → hypertension → antihypertensive (NSAID-induced hypertension treated with a new medication rather than discontinuing the NSAID); (2) Amlodipine → peripheral edema → diuretic (calcium channel blocker edema treated with a diuretic rather than switching antihypertensive class); (3) Cholinesterase inhibitor → urinary incontinence → anticholinergic (dementia drug causing incontinence treated with an anticholinergic that worsens cognition). Each cascade pattern encodes the temporal window, the typical symptom latency, and the evidence-based alternative intervention. Research indicates that up to 15% of new prescriptions in elderly patients represent prescribing cascades rather than treatment of new conditions.