30-day Patient Readmissions Predictive Model
The project connects healthcare feature engineering with predictive modeling through tuned XGBoost, Random Forest, and logistic regression classifiers for 30-day patient readmissions.
Connecting Patient History to Readmission Prediction
The 30-day readmission classification workflow connects healthcare data preparation with applied machine learning. It combines engineered features from ICD-coded conditions and discharge dispositions with patient-history features and classifiers built using XGBoost, Random Forest, and logistic regression.
Project focus: carrying healthcare data from domain-informed feature engineering into model development, hyperparameter tuning, and cross-validation.
Explore the implementation on GitHub
The Prediction Problem
The task was to classify whether a patient would be readmitted within 30 days. That target connects an encounter with a subsequent patient outcome and gives the modeling work a specific analytical purpose.
The challenge goes beyond choosing an algorithm. Clinical and administrative records need to become useful predictive inputs, and candidate models need an evaluation process. The project addresses those two responsibilities together: representing patient history and investigating how different classifiers use that representation.
From Data Foundation to Modeling Workflow
The project builds on the Healthcare FHIR Data Ingestion Framework. That foundation prepares healthcare data and patient-history features; this project carries those outputs into predictive modeling.
| Stage | Implementation | Deliverable |
|---|---|---|
| Data preparation | Use structured healthcare data from the ingestion framework | An analytical foundation for model development. |
| Feature engineering | Incorporate conditions, discharge dispositions, and patient history | Domain-informed predictive inputs. |
| Model development | Develop three classification approaches | XGBoost, Random Forest, and logistic regression models. |
| Evaluation | Perform hyperparameter tuning and cross-validation | Evaluation of candidate model configurations. |
This workflow connects source preparation to the prediction target. Feature choices and model choices are both part of the implementation.
Feature Engineering: Making Healthcare Context Usable
Clinical History
The modeling dataset incorporates custom features from ICD-coded conditions alongside chronic condition flags. These features bring recorded condition history into the modeling dataset.
The work connects domain understanding with implementation: identifying relevant healthcare information and representing it in a form that a classifier can consume.
Administrative and Encounter Context
The workflow combines features derived from discharge dispositions with prior admissions and encounter history trends from the ingestion framework. This combines clinical information with administrative and historical context.
| Feature Family | Context Represented |
|---|---|
| ICD-coded conditions | Recorded condition information. |
| Chronic condition flags | Persistent conditions in the patient history. |
| Discharge dispositions | Administrative context at discharge. |
| Prior admissions | Previous admission history. |
| Encounter trends | Patterns across the patient’s encounters. |
The feature engineering contribution is in bringing these different sources of context into a single classification task.
Model Development and Evaluation
Three Candidate Approaches
The project implements readmission classifiers using XGBoost, Random Forest, and logistic regression. These candidate approaches support model comparison and investigation of different configurations for the same predictive task.
Hyperparameter Tuning and Cross-Validation
The evaluation workflow uses hyperparameter tuning and cross-validation to assess model configurations and optimize predictive performance. This connects model development with systematic configuration evaluation.
The technical deliverables are the engineered features, candidate classifiers, and tuning and validation workflow.
Engineering Capabilities Demonstrated
The project brings together custom clinical and administrative features, three classification approaches, tuning, and cross-validation. It demonstrates the connections between:
- Healthcare domain context with feature design.
- Data engineering outputs with machine learning inputs.
- Model development with evaluation and configuration tuning.
The workflow spans data preparation and predictive modeling while keeping the readmission classification objective in view.
Deliverables and Analytical Value
The project delivers a readmission classification workflow that combines domain-informed features with multiple candidate models. It demonstrates how an engineered healthcare dataset can support an outcome-focused analytical task.
Together with the companion ingestion framework, it covers the full analytical workflow—from processing healthcare resources to developing and evaluating predictive models. Each project addresses a distinct part of that workflow, with the shared repository connecting the implementation.
Technical Stack
| Area | Technologies and Methods |
|---|---|
| Programming | Python |
| Modeling | XGBoost, Random Forest, logistic regression |
| Clinical Features | ICD-coded conditions, chronic condition flags |
| Administrative Features | Discharge dispositions, prior admissions, encounter history |
| Evaluation | Hyperparameter tuning, cross-validation |
| Upstream Foundation | PySpark-based FHIR ingestion and feature engineering |
Explore the Work
View the source repository for the implementation.
Read Healthcare FHIR Data Ingestion Framework to explore the data engineering behind the modeling workflow.