Clinical Analytics & 30-Day Diabetes Readmission Pipeline

An end-to-end healthcare data science and clinical analytics pipeline that identifies who returns for a 30-day hospital readmission among patients with diabetes.

This project integrates in-memory SQL extraction, machine learning classification, and clinical risk scoring to surface actionable patient cohorts for care coordinators and clinical teams.

Interactive Tableau Dashboard: View Live Dashboard on Tableau Public

---

Dataset & Clinical Source

This project utilizes the Diabetes 130-US Hospitals for Years 1999–2008 Dataset from the UCI Machine Learning Repository.

• Source: UCI Machine Learning Repository – Diabetes 130-US Hospitals
• Dataset Scope: 10 years of clinical care data representing 101,766 inpatient encounters across 130 US hospitals.
• Key Attributes: Demographics, admission/discharge clinical dispositions, length of stay, laboratory diagnostics (HbA1c, glucose), 23 diabetes medications, ICD-9 primary/secondary diagnoses, and 30-day readmission outcomes.

---

Overview & Clinical Objectives

30-day readmissions represent a critical quality metric and financial risk under hospital value-based purchasing programs. This pipeline solves two core operational challenges:

• Statistical Risk Prediction: Quantifies 30-day readmission risk via balanced Logistic Regression and extracts feature odds ratios ($\text{OR} = e^{\beta}$).
• Rule-Based Clinical Triage: Constructs an interpretable, point-based clinical risk index (Low, Moderate, High Priority) aligned with bedside decision rules.

---

Pipeline Architecture

  [ Raw Clinical Dataset (CSV) ]
                 │
                 ▼
  [ In-Memory SQLite Layer ]
    • ICD-9 primary diagnostic group mapping
    • Discharge & admission categorization
    • Exclusion criteria (mortality & hospice)
                 │
                 ▼
  [ Python Transformation & ML Engine ]
    ├─► Preprocessing: One-hot encoding & StandardScaler
    ├─► Logistic Regression (class_weight='balanced')
    ├─► Odds Ratio Calculation: exp(Beta)
    └─► Heuristic Point-Scoring & Risk Stratification
                 │
                 ▼
  [ Export Layer & Interactive BI ]
    ├─► feature_odds_ratios.csv
    └─► tableau_clinical_analytics.csv ──► Tableau Public Dashboard

---

Methodology & Logic

1. In-Memory SQL Cleaning & Clinical Mappings

Data ingestion and cleaning are executed in an embedded SQLite engine (:memory:) to apply clinical exclusions and standardize medical encodings:

• Exclusion Criteria: Patients discharged to hospice or expired during hospitalization (discharge_disposition_id IN (11, 13, 14, 19, 20, 21)) are excluded from readmission risk pools.
• ICD-9 Grouping: Primary diagnosis codes (diag_1) are categorized into standard clinical clusters (Circulatory, Respiratory, Diabetes, Digestive, Genitourinary, Musculoskeletal, Neoplasms, Injury/Poisoning).
• Discharge Destination: Mapped to standardized destination descriptions (Home, SNF, Rehab, Home Health).

2. Statistical Modeling & Odds Ratio Analysis

• Features include prior utilization (inpatient, emergency, outpatient), length_of_stay, medication counts, laboratory volume, diagnostic burden, and categorical clinical features.
• A balanced LogisticRegression model estimates patient-level predicted probabilities (Predicted_Readmit_Probability).
• Feature impact is quantified via exponentiated coefficients ($\text{Odds Ratio} = e^\beta$), categorizing features into Risk Increasing ($\text{OR} > 1.05$), Protective ($\text{OR} < 0.95$), or Neutral.

3. Clinical Heuristic Risk Score & Tiering

To complement probabilistic outputs with transparent bedside rules, a cumulative point score is calculated:

Clinical Feature	Threshold	Score Assigned
Prior Inpatient Visits	$\ge 2$ visits / $1$ visit	$+4$ pts / $+2$ pts
Prior Emergency Visits	$\ge 2$ visits / $1$ visit	$+2$ pts / $+1$ pt
Length of Stay	$\ge 7$ days / $\ge 4$ days	$+3$ pts / $+1$ pt
Medication Burden	$\ge 16$ total medications	$+2$ pts
Diagnostic Complexity	$\ge 8$ diagnoses	$+2$ pts
High-Risk Discharge	Discharged to SNF, Rehab, or Home Health	$+2$ pts
Glycemic Control	$\text{HbA1c} > 8%$	$+2$ pts

• Risk Tiers: $\le 3$ Points (Low Risk) | $4 - 7$ Points (Moderate Risk) | $> 7$ Points (High Risk Priority).

---

Tech Stack & Dependencies

• Language: Python 3.9+
• Data Ingestion & SQL: sqlite3, pandas
• Machine Learning & Preprocessing: scikit-learn (LogisticRegression, StandardScaler), numpy
• Business Intelligence: Tableau Public

---

Installation & Setup

1. Clone the repository:

   git clone [https://github.com/](https://github.com/)<your-username>/<repo-name>.git
   cd <repo-name>

2. Install dependencies:

   pip install pandas numpy scikit-learn

3. Run the pipeline: Place the raw dataset diabetic_data.csv in the root directory and run:

   Workflow.py

---

Deliverables & Dashboard

• tableau_clinical_analytics.csv: Master enriched dataset containing demographic data, utilization flags, predicted probabilities, and clinical risk tiers.
• feature_odds_ratios.csv: Statistical odds ratios for all model covariates.
• Interactive Dashboard: Live visual interface for clinical cohort drill-down, risk tier distribution, and feature impact analysis.