All projects
Machine Learning/Healthcare analysis /Project 04 of 10 M.S. coursework

Hospital Readmission Analysis

A paired supervised and unsupervised analysis of hospital outcomes, combining discharge-time risk stratification with patient clustering and PCA.

Python Healthcare Risk stratification Feature timing Clustering PCA M.S. coursework

Business impact

Combines a discharge-time readmission classifier with patient segments to inform questions about follow-up and hospital service use. The study did not test an intervention.

0.999ROC-AUC at discharge
97.9%Held-out classifier accuracy
k = 3Patient segments

Business question

How well can hospital records classify readmission, and what patient groups emerge from clustering?

Approach

How the work was done

  • Compared 216 Random Forest parameter combinations with five-fold cross-validation, then evaluated the selected forest on a held-out test set.
  • Used the Elbow Method to select three K-Means segments, then applied PCA to inspect cluster separation and profiles.

Results

What the analysis shows

  • The full discharge-time model achieved 97.9% accuracy, 96.3% precision, 98.0% recall, 0.971 F1, and 0.999 ROC-AUC.
  • The separate clustering notebook uses 14 features and excludes readmission. One group averaged 9 days and $3,212 in charges; another averaged 60 days and $7,401. The oldest group averaged 74 years, and 95% reported high blood pressure. These profiles do not establish future readmission risk.

From the project

Model evaluation and patient groups

ROC curves with and without completed length of stay and a full-model confusion matrix showing 929 true negatives, 21 false positives, 11 false negatives and 539 true positives.

How performance depends on feature timing

The same held-out records give ROC-AUC 0.999 with completed length of stay and 0.492 without it. The confusion matrix describes the full model at the evaluated threshold.

View the source Open full size
Saved elbow plot from the current 14-feature clustering notebook, with within-cluster variation decreasing from one to ten groups.

Choosing the number of patient groups

The early reductions in within-group variation informed the choice of three clusters. These groups describe similar records; they do not establish future readmission risk.

View the source Open full size

Additional validation

Feature timing at discharge

Initial_days records completed length of stay, available at discharge. A sensitivity check removes that feature and repeats the selected-model evaluation. ROC-AUC falls from 0.999 to 0.492, near chance. This limits the full model to discharge-time analysis; decisions at admission need a model validated on information available then.

Business relevance

What this makes easier to decide

The classifier provides a way to rank records for discharge-time review, while cluster profiles help explore differences in service use. Neither analysis establishes that follow-up will reduce readmissions.