Review: EHR-Linked AI Scores Guide VTE Prophylaxis and Monitoring

• Across 23 studies, AI models predicted inpatient VTE risk with test AUCs 0.740–0.990 and generally outperformed traditional scores; 16 models had external validation (validation AUC 0.740–0.980).

• For pharmacy teams, validated, EHR-linked AI VTE risk scores can drive risk‑stratified workflows—prioritizing prophylaxis verification and anticoagulation monitoring for patients with age ≥70, obesity, prior VTE, abnormal coagulation labs, surgical history, decreased mobility, or CVC/PICC.

• VTE is a common, preventable inpatient harm; traditional scores (Caprini, Padua) use static, limited variables and show only modest accuracy (AUC ~0.60–0.75), contributing to both over‑ and under‑prophylaxis.

• Hospital pharmacy needs timely, EHR‑ready risk signals to coordinate prophylaxis and monitoring. AI can fuse labs, comorbidities, procedures, catheter status, and clinical text to produce dynamic risk predictions, but real‑world use is constrained by interpretability, data standardization, cross‑site generalizability, and EHR integration.

• Reliable, validated risk stratification is foundational for dependable clinical decision support that targets VTE prevention and anticoagulation management.

• Scoping review of Chinese and English databases through 10 March 2025; screened 3,318 records and included 23 adult studies spanning hospitalized, cancer, orthopedic, surgical, stroke, and trauma cohorts.

• Standardized data extraction captured model types and inputs (random forest, gradient boosting, natural language processing, neural networks, support vector machines, ensembles), study design, sample source, and performance metrics; two reviewers double‑extracted data (Cohen’s κ = 0.84).

• Recorded comparisons to Caprini/Padua scores, internal versus external validation, and noted that most source data were retrospective, single‑center EHRs or registries.

• AI models across studies achieved test‑set AUCs 0.740–0.990; 16 of 23 models had external validation with validation AUCs 0.740–0.980, indicating broadly reproducible discrimination where validated.

• Tree‑based methods dominated (10 random forest, 5 gradient boosting); random forests commonly reported AUC >0.85, NLP approaches reached up to ~0.97–0.98; sensitivity ranged ~0.29–1.00 and specificity ~0.59–0.998 across studies.

• Consistent top predictors (reported in ≥10 studies) included age ≥70, obesity (BMI ≥30 kg/m2), prior VTE, abnormal coagulation labs (D‑dimer, fibrinogen), surgical history, decreased mobility, and CVC/PICC presence.

• Pharmacy‑relevant takeaway: validated, EHR‑integrated AI risk scores—driven by richer inputs (labs, comorbidities, procedures, and unstructured note text), modern algorithms, and external validation—could help pharmacists prioritize prophylaxis and target anticoagulation monitoring, reducing unnecessary prophylaxis in low‑risk patients.

• Embed an EHR VTE risk score that populates a pharmacy worklist and triggers tiered actions: high risk—verify prophylaxis, monitor anticoagulation, reassess catheter necessity; rising risk—trend D‑dimer/fibrinogen and prompt mobility interventions.

• Prepare data and interfaces: standardize lab units/codes, map surgery/ventilation/length‑of‑stay and catheter fields, include usable clinical note text, and deliver real‑time EHR interfaces (CDS banners, worklists, dashboards) that track alerts and actions.

• Implementation pathway: perform local validation on historical cohorts, pilot on selected units, calibrate alert thresholds to match review capacity, and train staff using concise feature summaries (age ≥70, obesity, prior VTE, abnormal labs, surgery, CVC/PICC).

• Operational principle: treat the model as a detection tool—maintain human‑in‑the‑loop governance with override rules, clear escalation paths, and ongoing performance monitoring across services and patient groups.

Strengths:

• Rigorous bilingual search across Chinese and English databases, with standardized double extraction and high inter‑rater agreement (Cohen’s κ = 0.84).

• Clear methodological mapping—cataloged algorithms, inputs, validation types, and recurrent risk factors across diverse cohorts—supporting reproducibility and practical implementation planning.

Limitations:

• Evidence base is skewed toward retrospective, single‑center EHR studies; seven models lacked external validation and cross‑site generalizability remains uncertain given data standardization challenges.

• No pooled meta‑analysis or prospective, real‑time deployment outcome studies; limited interpretability of many tree/ensemble models and the technical requirements for EHR integration constrain near‑term clinical adoption.

Validated AI VTE risk models provide accurate, EHR‑ready stratification that can enable pharmacist‑led, risk‑targeted prophylaxis and anticoagulation monitoring—but require local validation, data standardization, EHR integration pilots, and governance before scale‑ups.