Patient Mortality & Readmission Prediction

Problem Statement

Predicting patient mortality and readmission risk is critical for hospital resource allocation and preventative care. Electronic health records contain both structured data (lab results, vitals, demographics) and unstructured clinical notes (physician observations, discharge summaries). Most ML models use only one modality, leaving valuable signal unused.

Technical Approach

Multimodal Fusion Architecture

The proposed design uses a late-fusion architecture to combine predictions from models operating on different data modalities:

• Structured Data Model: An XGBoost workflow using features such as lab values, vital signs, medication history, and demographic indicators, with hyperparameter tuning through AWS SageMaker.
• Unstructured Text Model: A ClinicalBERT workflow that extracts semantic features from notes, including symptoms, diagnoses, and treatment plans.
• Fusion Layer: A weighted ensemble that combines XGBoost probability scores with BERT representations through a logistic-regression meta-learner.

Reference AWS Architecture

A possible cloud architecture for the prototype includes:

• AWS Glue: ETL for missing-value handling, normalization, and feature engineering
• Amazon Athena: SQL-based exploratory analysis on data stored in S3
• AWS Lambda: Event-driven orchestration for prototype inference tasks
• Amazon SageMaker: Model training, tuning, and evaluation
• Amazon Bedrock: Optional exploration of human-readable model summaries

Evaluation Plan

• Compare the structured, text, and fused models with the same held-out evaluation protocol
• Report discrimination, calibration, and inference-time measurements only when they are reproducible
• Use public, synthetic, or appropriately authorized de-identified data for experimentation
• Inspect feature importance and error patterns while avoiding claims of clinical validity
• This page describes a personal or academic case study, not a clinical deployment or client engagement

Tech Stack