CareIQ — Patient Triage & Readmission Predictor

The Challenge

Unplanned 30-day hospital readmissions cost the U.S. healthcare system $26 billion annually, and CMS penalizes hospitals up to 3% of Medicare reimbursements for excessive readmission rates. The Healthcare Analytics Group's client network — a 3-hospital system — had a 30-day readmission rate of 18.4%, compared to the national benchmark of 15.1%. Discharge planning was based on clinical intuition and standard protocol checklists — not patient-specific risk stratification. High-risk patients received the same discharge instructions as low-risk ones. Follow-up calls were made by an overwhelmed nursing team working from paper lists, with no priority ordering. The result: the patients who most needed post-discharge intervention were the least likely to receive timely outreach. The challenge was not data availability — the network had 2M+ historical patient records — but the absence of a system to translate that data into actionable, real-time risk scores embedded directly into nursing workflows.

Our Custom Solution

CareIQ is a predictive ML platform that generates a 30-day readmission risk score for every patient at the moment of discharge. The core model is an XGBoost gradient-boosted classifier trained on 2.1M de-identified patient records spanning 6 years, incorporating 47 clinical and social determinant features: primary diagnosis, comorbidity index, length of stay, medication count, prior admissions in 12 months, discharge disposition, insurance type, zip code poverty index, and 38 additional variables. SHAP (SHapley Additive exPlanations) values render the model interpretable — nurses see not just a risk score but the top 5 contributing factors for each patient ("High risk due to: 3 prior admissions, CHF diagnosis, lives alone, no follow-up appointment scheduled"). AWS SageMaker hosts the model endpoint for real-time scoring at discharge. High-risk patients (score ≥ 0.72) are automatically queued in a Twilio-powered outreach system that calls patients at 24h, 72h, and 7-day post-discharge intervals with structured check-in scripts. The React nursing dashboard surfaces the risk queue with one-click call logging and exception escalation to the care coordination team.

System Architecture & Technologies

• XGBoost classifier trained on 2.1M patient records with 47 clinical + SDOH features
• SHAP explainability layer providing per-prediction factor attribution for clinical transparency
• AWS SageMaker real-time inference endpoint with p99 latency < 180ms
• FastAPI score ingestion API integrated into 3 hospital EHR discharge workflows
• Twilio Programmable Voice for automated multi-touch post-discharge outreach calls
• PostgreSQL patient tracking database with 90-day rolling readmission outcome data
• React nursing dashboard with risk stratification queue, call logging, and escalation workflows
• Docker-based model retraining pipeline running monthly on new discharge outcome data

Technologies: XGBoost • SHAP • AWS SageMaker • Twilio • React • PostgreSQL • FastAPI • Docker

Measurable Results & Outcomes

• 41%: Reduction in 30-day readmission rate (18.4% → 10.9%)
• 0.91: AUC-ROC score — top-decile predictive accuracy for readmission models
• 2.1M: Patient records used for model training and validation
• $3.8M: Estimated annual CMS penalty avoidance across 3 hospitals
• 3: Hospital systems fully integrated with live EHR discharge workflows
• 76%: Of high-risk patients reached within 24 hours of discharge via Twilio

Client Testimonial

"We went from readmission rates that were costing us $3M+ in CMS penalties to being in the top quartile nationally. The SHAP explanations were key — nurses actually trust the scores because they understand why."

Explore More Case Studies

• SmartSite Vision — Construction Site Safety AI (73% incident reduction)
• MediFlow — Clinical Document Intelligence (4h+ saved per physician/day)
• BuildBot — Construction Project Manager AI (31% on-time improvement)
• OpsGenie AI — Enterprise Operations Hub (60% admin time saved)