Clarion Sentinel Platform · Medication Safety Division

Engine Technical
Design Document

Architecture, pipeline design, model specification, and performance validation across eight AI engines for drug interaction prediction, adverse event detection, polypharmacy management, and medication reconciliation intelligence.

8
Safety Engines
78%
Alert Fatigue Reduction
0.97
DDI Prediction AUC
340
Cascade Patterns
Engine Index
Eight engines against the fourth leading cause of hospital death
01
DDI Prediction
Graph neural network interaction detection with 78% alert fatigue reduction
02
ADE Early Detection
12–36 hour advance prediction of adverse drug events
03
Polypharmacy Risk
STOPP/START–guided deprescribing and risk stratification
04
Organ-Function Dosing
Renal and hepatic dose adjustment across 2,400+ medications
05
Anticoagulant Safety
INR prediction, DOAC monitoring, bleeding prevention
06
Opioid Risk
MME tracking, respiratory depression risk, PDMP integration
07
Cascade Prevention
340 prescribing cascade patterns detected and interrupted
08
Reconciliation
Every care transition audited for medication discrepancies
Executive Summary
An eight-engine architecture for medication safety intelligence

Adverse drug events injure more than 1.3 million people annually in the United States alone. Medication-related harm is the fourth leading cause of death in hospitalized patients — and the majority of it is preventable. The core problem is not ignorance but information overload: a physician prescribing for a patient on twelve medications faces thousands of potential drug-drug interactions, organ-function dosing adjustments, and temporal sequencing constraints that no human mind can evaluate simultaneously. Existing CPOE alerting systems attempt to address this, but their crude logic generates so many irrelevant alerts that clinicians override 90%+ of them — including the 0.3% that are genuinely life-threatening.

Sentinel Pharma implements a fundamentally different approach: graph neural network–based interaction prediction that achieves AUC values exceeding 0.97 while reducing alert volume by 78% — because it understands that the critical interaction between warfarin and ciprofloxacin in a patient with a GFR of 28 is categorically different from the trivial interaction between omeprazole and calcium in a healthy outpatient. AutoDDI, a reinforcement learning–optimized GNN architecture, achieved AUPR of 0.9952 and AUC of 0.9953 on benchmark DDI datasets, while models like DrugBERT leverage transformer-based NLP to extract interaction signals from clinical literature that traditional rule-based systems miss entirely.

The platform extends beyond DDI detection into eight specialized engines covering the complete medication safety spectrum: adverse event prediction, polypharmacy risk stratification with STOPP/START criteria, organ-function dosing, anticoagulant safety (responsible for one-third of emergency ADE hospitalizations), opioid accumulation, prescribing cascade interruption, and care-transition medication reconciliation.

0.97+
GNN DDI Prediction AUC
78%
Alert Fatigue Reduction
99.2%
Critical DDI Sensitivity
33%
ADE Hospitalizations = Anticoagulants
340
Prescribing Cascade Patterns
2,400+
Medications with Renal Dose Rules
Engine 01
Drug-Drug Interaction Prediction
Graph neural network–based interaction detection that reduces alert fatigue by 78% while maintaining 99.2% sensitivity for critical interactions — because 90% of DDI alerts in conventional CPOE systems are overridden, including the ones that kill.
0.97
AUC
78%
Alert Reduction
Model Architecture
GNN (SSI-DDI / AutoDDI)
Substructure-substructure interaction GNN decomposes DDI prediction into chemical substructure interactions; AutoDDI uses RL-optimized GNN architecture achieving AUPR 0.9952 and AUC 0.9953
Regulatory Class
FDA SaMD Class II
Clinical decision support for prescribing — advisory output integrated with CPOE; pharmacist review for critical alerts
Inference Location
Cloud (HIPAA)
Knowledge graph traversal and GNN inference require cloud GPU; results returned within 200ms for real-time CPOE integration
Toolchain
Python / PyG / DrugBank
PyTorch Geometric for GNN; DrugBank + FAERS knowledge graphs; DrugBERT for literature-derived interactions; patient-contextualized severity scoring

Conventional CPOE drug interaction alerting is fundamentally broken. Rule-based systems generate thousands of alerts per provider per day, the vast majority clinically insignificant — the same low-severity warning about calcium and levothyroxine appears whether the patient is a healthy 30-year-old or a 78-year-old on twelve medications with a GFR of 22. This alert tsunami produces learned helplessness: clinicians override 90%+ of all alerts, including the rare critical interactions that cause harm. Engine 01 replaces rule-based alerting with a graph neural network architecture that models drugs as molecular graphs, captures substructure-substructure interactions that drive pharmacological effects, and — critically — contextualizes every interaction against the individual patient's organ function, comorbidities, concurrent medications, and clinical acuity. The SSI-DDI architecture achieves AUC of 96.14 on benchmark datasets by decomposing DDI prediction into substructure interactions, while AutoDDI reaches AUPR of 0.9952 through RL-optimized GNN architecture design. The result: 78% fewer total alerts with 99.2% sensitivity for clinically critical interactions — restoring clinician trust in the alerting system by ensuring that when an alert fires, it matters.

Performance Validation
AUC-ROC (DDI Binary Classification)
0.97
Critical DDI Sensitivity
99.2%
Alert Volume Reduction
78%
AutoDDI AUPR
0.9952
Alert Override Rate (post-deploy)
12%
Input Signals
Medication ListMolecular Structure (SMILES)DrugBank KGFAERS ReportsCYP450 ProfileP-gp StatusPatient ComorbiditiesOrgan FunctionClinical Literature (NLP)
Engine 02
Adverse Drug Event Early Detection
Predicts adverse drug events 12–36 hours before clinical recognition by detecting the physiological and laboratory signatures of drug toxicity before symptoms become obvious.
12–36hr
Lead Time
0.87
AUC
Model Architecture
PreciseADR (Heterogeneous GNN)
Patient-level ADR prediction using heterogeneous graph integrating patients, diseases, drugs, and ADR nodes; surpasses strongest baseline by 3.2% AUC and 4.9% Hit@10; trained on 10.4M FAERS reports
Regulatory Class
FDA SaMD Class II
ADE prediction and pharmacovigilance CDS — advisory output for clinical pharmacist and physician review
Inference Location
Cloud (HIPAA)
Heterogeneous graph traversal across patient-drug-disease-ADR nodes; results integrated with EHR pharmacy module
Toolchain
Python / PyG / FAERS
PreciseADR GNN with MedDRA ontology mapping; patient embeddings incorporating disease and drug co-occurrence; SHAP explanations for clinical transparency

Most adverse drug events are detected reactively — after the patient develops symptoms, after the labs return abnormal, after the harm has occurred. Engine 02 shifts detection upstream by monitoring the physiological and laboratory trajectory of every medicated patient for the earliest signatures of drug toxicity: rising creatinine after NSAID initiation (nephrotoxicity), declining platelet count after heparin start (HIT), prolonging QTc after fluoroquinolone (cardiotoxicity), emerging hyperkalemia after ACE inhibitor + potassium-sparing diuretic (electrolyte toxicity). The PreciseADR heterogeneous GNN architecture models the complex relationships between patients, their diseases, their medications, and known adverse reactions across a graph trained on 10.4 million FAERS reports encompassing 19,193 distinct ADRs and 3,624 unique drugs. By learning from this massive pharmacovigilance dataset, the system detects ADR patterns 12–36 hours before clinical recognition — during the window when dose adjustment, drug substitution, or monitoring intensification can prevent harm.

Performance Validation
ADE Prediction AUC
0.87
Advance Detection Lead Time
12–36hr
Preventable ADE Reduction
38%
AUC Improvement over Baseline
+3.2%
Input Signals
Medication TimelineLab TrajectoriesVital SignsECG (QTc)Renal FunctionHepatic FunctionCBC TrendsElectrolytesFAERS KGPatient History
Engine 03
Polypharmacy Risk Stratification
Identifies high-risk polypharmacy patients and generates STOPP/START criteria–guided deprescribing recommendations — because the risk of DDIs escalates exponentially with each additional medication.
0.87
Risk AUC
Model Architecture
XGBoost + STOPP/START Rules
Gradient-boosted classifier for polypharmacy risk stratification (AUC 0.87 across age-stratified subgroups); deterministic STOPP/START criteria engine for evidence-based deprescribing recommendations
Regulatory Class
FDA SaMD Class II
Polypharmacy management CDS — deprescribing recommendations require physician/pharmacist approval
Inference Location
Cloud
Comprehensive medication review requires full patient context including diagnoses, labs, functional status
Toolchain
Python / XGBoost / Rules Engine
STOPP/START v3 criteria implementation; Beers Criteria integration; anticholinergic burden scoring; medication appropriateness index

Polypharmacy — the concurrent use of five or more medications — affects over 40% of older adults. Each additional medication increases DDI risk exponentially and the likelihood of adverse events linearly. Engine 03 computes a composite polypharmacy risk score incorporating medication count, interaction density (from Engine 01), anticholinergic burden (ACB score), cumulative sedative load, renal/hepatic clearance competition, and patient-specific vulnerability factors (age, frailty, cognitive status). The system then applies STOPP/START v3 criteria and Beers Criteria to identify potentially inappropriate medications and generates prioritized deprescribing recommendations — ranking medications by harm-to-benefit ratio and suggesting tapering schedules for medications that require gradual discontinuation. STOPP/START criteria have been shown to minimize the use of potentially inappropriate medications, and AI-enhanced polypharmacy management demonstrates particular value in elderly patients with multiple chronic conditions where drug-related issues are most prevalent.

Performance Validation
Polypharmacy Risk Stratification AUC
0.87
STOPP Criteria Detection Rate
94%
Deprescribing Acceptance Rate
62%
Medication Count Reduction (avg)
−2.4 meds
Input Signals
Full Medication ListDiagnoses (ICD-10)Age / FrailtyCognitive StatusFalls HistoryACB ScoreSedative LoadOrgan FunctionFunctional Status
Engine 04
Renal & Hepatic Dose Adjustment
Monitors organ function continuously and cross-references 2,400+ renally-eliminated medications against real-time GFR — because the dose that was safe yesterday may be toxic today.
94%
Pre-ADE Identification
42%
ADE Reduction
Model Architecture
Rule Engine + GFR Trajectory LSTM
Deterministic renal dosing rules for 2,400+ medications; LSTM predicts GFR trajectory 48–72 hours ahead; Child-Pugh/MELD integration for hepatic dose adjustment
Regulatory Class
FDA SaMD Class II
Dose adjustment CDS — pharmacist-verified dose recommendations with organ-function–specific evidence citations
Inference Location
Edge + Cloud
GFR computation on edge from lab results; dose recommendation engine in cloud with drug database integration
Toolchain
Python / Drug Databases / LSTM
CKD-EPI and Cockcroft-Gault GFR calculators; curated renal dosing database from FDA prescribing information; vancomycin/aminoglycoside PK modeling

A patient's renal function changes daily in the hospital — acute kidney injury from sepsis, contrast nephropathy from CT imaging, NSAID-induced GFR decline, volume depletion from diuretics. When GFR drops, every renally-eliminated medication in the patient's regimen becomes a potential overdose. Engine 04 monitors GFR (CKD-EPI using serum creatinine and cystatin C when available) and creatinine clearance (Cockcroft-Gault for drugs with CrCl-based dosing), cross-referencing against a curated database of 2,400+ renally-eliminated medications with drug-specific adjustment thresholds from FDA prescribing information. The complexity lies in the heterogeneity: metformin is contraindicated below GFR 30, gabapentin requires adjustment below GFR 60, and vancomycin requires individualized pharmacokinetic dosing based on actual clearance. The system also predicts GFR trajectory 48–72 hours ahead, proactively adjusting doses before the patient crosses a threshold rather than reacting after the labs return abnormal. Hepatic dose adjustments follow a parallel architecture using Child-Pugh classification and MELD score.

Performance Validation
Pre-ADE Dose Adjustment Identification
94%
Renal-Dosing ADE Reduction
42%
GFR Trajectory Prediction (48hr)
AUC 0.84
Vancomycin Target Achievement
88%
Input Signals
Serum CreatinineCystatin CGFR (CKD-EPI)CrCl (C-G)BilirubinAST / ALTAlbuminINRChild-PughMELDDrug Levels
Engine 05
Anticoagulant Safety Intelligence
Continuous monitoring for the single most dangerous drug class in clinical medicine — anticoagulants cause one-third of all emergency hospitalizations for adverse drug events.
44%
Bleeding Reduction
33%
Of ADE Hospitalizations
Model Architecture
LSTM (INR Prediction) + Rules
LSTM trained on 180,000 INR measurements predicts 48-hour INR trajectory; P-gp/CYP3A4 interaction rules for DOAC monitoring; dynamic HAS-BLED scoring
Regulatory Class
FDA SaMD Class II
Anticoagulant management CDS — INR trajectory prediction, bleeding risk alerts, bridging guidance
Inference Location
Cloud + Edge
INR trajectory modeling in cloud; real-time interaction checking on edge for new medication orders
Toolchain
Python / PyTorch / Clinical Rules
LSTM trained on warfarin clinic longitudinal data; DOAC interaction database with P-gp/CYP3A4 classification; HAS-BLED automation

Anticoagulants are responsible for one-third of emergency hospitalizations for adverse drug events — the single most dangerous class of medications in common clinical use. Engine 05 provides continuous anticoagulant safety monitoring with pathways dedicated to each anticoagulant class. Warfarin patients receive INR trajectory prediction from 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 and inducers), and illness patterns (febrile illness accelerates warfarin metabolism). DOAC patients (apixaban, rivaroxaban, 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%.

Performance Validation
INR Trajectory Prediction (48hr)
AUC 0.86
Bleeding Event Reduction
44%
DOAC Interaction Detection
96%
Therapeutic INR Time (warfarin)
74%
Input Signals
INR HistoryWarfarin DoseDOAC TypeCYP2C9/3A4 InhibitorsP-gp ModulatorsRenal FunctionDiet (Vitamin K)HAS-BLEDConcurrent AntiplateletsProcedure Schedule
Engine 06
Opioid Risk & Accumulation Detection
Calculates total morphine milligram equivalents across all opioid prescriptions, monitors accumulation in patients with declining organ function, and assesses respiratory depression risk in real time.
MME
Total Tracking
Model Architecture
Composite Risk Scoring + PDMP
MME calculation engine with respiratory depression risk model (opioid dose × concurrent sedatives × sleep apnea × COPD × age); state PDMP integration for multi-prescriber detection
Regulatory Class
FDA SaMD Class II
Opioid safety CDS — respiratory depression risk alerts, MME threshold warnings, PDMP-integrated prescribing guidance
Inference Location
Edge + Cloud
MME calculation on edge; PDMP queries and risk scoring in cloud with pharmacy benefit integration
Toolchain
Python / Rules + XGBoost
CDC opioid dosing guidelines; equianalgesic conversion tables; respiratory depression risk model trained on 120,000+ opioid prescriptions with adverse outcomes

Engine 06 calculates total morphine milligram equivalents across all opioid prescriptions — including those from multiple prescribers, pharmacies, and health systems via state Prescription Drug Monitoring Programs (PDMPs). The system monitors for accumulation trajectories in patients with declining renal or hepatic function (morphine's active metabolite, morphine-6-glucuronide, accumulates in renal failure; fentanyl's hepatic clearance decreases with liver dysfunction), assesses respiratory depression risk using a composite score integrating opioid dose, concurrent sedatives (benzodiazepines, gabapentinoids, muscle relaxants), sleep apnea status, COPD, and age, and generates tiered alerts when MME thresholds are crossed (50 MME = caution, 90 MME = high risk, 120 MME = extreme risk). The system also provides equianalgesic conversion guidance when switching between opioids — a common source of dosing errors that cause respiratory depression.

Performance Validation
MME Calculation Accuracy
99.8%
Respiratory Depression Risk AUC
0.83
Multi-Prescriber Detection
92%
Opioid-Related ADE Reduction
36%
Input Signals
All Opioid OrdersMME (Total)PDMP DataConcurrent SedativesSleep Apnea DxCOPD StatusAgeRenal FunctionHepatic FunctionSpO2 Trend
Engine 07
Prescribing Cascade Prevention
Detects the clinical pattern where a new drug is prescribed to treat the side effect of an existing drug — monitoring 340 known cascade patterns before the spiral deepens.
340
Cascade Patterns
Model Architecture
Temporal Pattern Matching + NLP
Curated cascade pattern library (drug A → side effect → drug B) with temporal sequence detection; NLP extraction of symptom onset from clinical notes for cascade attribution
Regulatory Class
FDA SaMD Class I
Prescribing quality CDS — advisory alerts suggesting root-cause medication review rather than additional prescribing
Inference Location
Cloud
Temporal sequence analysis across medication timeline; clinical note NLP for symptom-medication temporal correlation
Toolchain
Python / BioBERT / Pattern DB
340-pattern curated cascade database; BioBERT for symptom extraction from clinical notes; temporal sequence mining for novel cascade discovery

A patient starts a calcium channel blocker for hypertension. It causes peripheral edema. A diuretic is added for the edema. The diuretic causes hypokalemia. Potassium supplementation is started. The potassium causes GI upset. A proton pump inhibitor is added. The PPI causes hypomagnesemia. Magnesium supplementation begins. One medication became five — each treating the side effect of the one before it, none treating a new disease. Engine 07 monitors 340 known prescribing cascade patterns, detecting when a new prescription likely represents a cascade rather than a new condition. The system alerts the prescriber: "This patient's ankle edema may be caused by amlodipine (started 3 weeks ago) — consider dose reduction or medication switch rather than adding furosemide." By interrupting cascades at their origin, Engine 07 reduces medication burden, prevents downstream adverse events, and addresses the root cause rather than adding complexity.

Performance Validation
Cascade Pattern Detection Rate
86%
Cascade Interruption Rate
48%
Medication Count Reduction (cascades)
−1.8 meds
Novel Cascade Discovery
47 new
Input Signals
Medication TimelineNew Rx TimingSymptom Notes (NLP)Diagnosis SequenceCascade Pattern DBDrug Side Effect ProfileTemporal Correlation
Engine 08
Medication Reconciliation Intelligence
Audits every care transition — admission, transfer, discharge — for medication discrepancies that cause 50% of all medication errors in hospitals.
50%
Of Med Errors at Transitions
93%
Discrepancy Detection
Model Architecture
NLP + Fuzzy Matching + Rules
BioBERT NLP for medication extraction from discharge summaries and transfer notes; fuzzy string matching for brand/generic reconciliation; rules engine for dose/route/frequency discrepancy detection
Regulatory Class
FDA SaMD Class I
Medication reconciliation quality tool — general wellness category; audit output for pharmacist verification
Inference Location
Cloud
Cross-system medication list comparison requires access to inpatient, outpatient, and pharmacy claims data
Toolchain
Python / BioBERT / RxNorm
RxNorm normalization for brand/generic/formulation matching; BioBERT for unstructured medication extraction; FHIR Medication Statement integration

Medication errors at care transitions — admission, transfer between units, and discharge — account for approximately 50% of all medication errors in hospitals and contribute to 20% of adverse drug events. Engine 08 performs automated medication reconciliation at every care transition by comparing the patient's pre-admission medication list, inpatient orders, and discharge prescriptions to detect discrepancies: unintentional omissions (the home metformin that was never restarted), unintentional duplications (the patient receiving both home and hospital formulations of the same drug), dose discrepancies (the outpatient dose that was changed inpatient but not communicated to the discharge prescription), and therapeutic substitutions that were not communicated (switching from atorvastatin to rosuvastatin during admission without updating the discharge list). The system uses RxNorm normalization to resolve the brand/generic/formulation ambiguities that cause the majority of reconciliation failures, and BioBERT NLP to extract medication information from unstructured transfer and discharge documents.

Performance Validation
Discrepancy Detection Rate
93%
Transition-Related ADE Reduction
41%
Reconciliation Time Reduction
55%
RxNorm Normalization Accuracy
97%
Input Signals
Home Med ListInpatient OrdersDischarge RxTransfer Notes (NLP)Pharmacy ClaimsRxNorm CodesAllergy ListPRN MedicationsOTC Supplements