CMS Hospital Readmission Risk Analysis

A complete end-to-end healthcare analytics solution using CMS hospital readmissions data — applying Python for data cleansing and predictive modelling, and delivering an interactive Power BI dashboard presenting 30-day readmission risk stratification, patient cohort trends, and cost impact indicators to support clinical decision-making.

---

Project Overview

Item	Detail
Dataset	5,000 CMS-style patient encounters
Period	Jan 2023 – Feb 2024
Hospitals	6 facilities
Diagnoses	10 CMS conditions
Champion model	Gradient Boosting (GBM)
Model AUC-ROC	0.627
Readmission rate	64.3%
Critical risk patients	3,017 (60.3%)
Total dataset cost	$147.9M

---

Project Structure

cms-readmission-risk-analysis/

• Notebook_01_Data_Collection.ipynb
• Notebook_02_Data_Cleansing.ipynb
• Notebook_03_Feature_Engineering.ipynb
• Notebook_04_Modelling.ipynb
• CMS_Readmission_Dashboard.pbix
• requirements.txt
• cms_patients_raw.csv
• fact_patients_scored.csv
• agg_monthly_kpis.csv
• agg_diagnosis.csv
• agg_hospital.csv
• agg_risk_tiers.csv
• patient_watchlist.csv
• feature_importance.csv
• dim_date.csv
• cleansing_report.json
• model_performance.json

---

Pipeline Steps

Step 1 — Data Collection

Generates a realistic 5,000-record CMS-style inpatient dataset with clinically grounded distributions. Injects 5% dirty records to demonstrate the cleansing step.

Step 2 — Data Cleansing

Applies 15 data quality rules across five severity levels:

Rule	Issue	Severity
R01	Duplicate patient IDs	CRITICAL
R02	Age outside 18–95	HIGH
R03	LOS outside 1–30 days	HIGH
R04	LOS vs date diff mismatch	MEDIUM
R05	BMI outside 10–60	HIGH
R06	Cost zero or negative	HIGH
R07	Cost extreme outlier (>3×IQR)	MEDIUM
R08	LOS extreme outlier (>20 days)	LOW
R09–R15	Formatting, flags, audit columns	INFO

Step 3 — Feature Engineering

Transforms 36 clean columns into 72 model-ready features:

• Temporal features (9)
• Demographic bands (7)
• Binary clinical flags (13)
• Composite risk scores (4)
• Encoded categoricals (15)

Step 4 — Predictive Modelling

Trains and evaluates three models:

Model	AUC-ROC	Sensitivity	Specificity
Logistic Regression	0.653	87.9%	28.0%
Random Forest	0.654	93.9%	11.5%
Gradient Boosting ★	0.627	83.7%	31.4%

---

Power BI Dashboard (6 Pages)

Page	Content
Executive Summary	KPI overview and navigation
Clinical Overview	Risk tiers, diagnosis rates, monthly trends
Risk Stratification	Patient watchlist with slicers
Cohort Trends	Age, payer, LOS trend analysis
Cost Impact	Cost by tier, diagnosis scatter
Hospital Benchmarking	Hospital comparison and feature importance

---

DAX Measures

Measure	Purpose
Total Patients	Count of all patient encounters
Readmissions Count	Count of readmitted patients
Readmission Rate	Readmissions / Total patients
Critical Patients	Count of Critical-tier patients
Total Cost	Sum of all encounter costs
Avg Risk Score	Mean model risk score
Readmission Cost	Total cost of readmitted patients
Avg LOS	Average length of stay

---

Technologies Used

• Python 3.10
• pandas, numpy, scikit-learn
• Jupyter Notebook
• Power BI Desktop
• DAX (Data Analysis Expressions)

---

How to Run

1. Clone the repository
2. Install dependencies: