Devam Shah

← All projects

ML & Deep Learning

ReadmitRadar: Hospital Readmission Network

A neural network built from first principles to predict 30-day hospital readmission, with manual forward propagation, a loss-function bake-off and five documented regularization trials.

62.6%test accuracy with a narrower 64 → 32 → 16 network

View the code on GitHub ↗
Stack: Python, TensorFlow/Keras, pandas, scikit-learn License: MIT

Live demo

A working version of this project will run right here.

Coming soon

Plan: Compact model exported to TensorFlow.js so visitors can enter patient features and see a readmission probability in the browser, framed honestly as an educational 62.6%-accuracy model.

TODO: build and embed the demo

Key findings

  • Built on 5,000 clinical records with 14 features.
  • Five documented regularization trials were used to diagnose overfitting instead of just patching it.
  • The best model combines L1+L2, batch normalization, dropout, early stopping and learning-rate scheduling.

How it works

  • Manual forward propagation to understand the mechanics before using the framework.
  • A loss-function bake-off to choose the training objective.
  • Write-ups: Medium, "Your Dropout Layer Isn't Fixing Overfitting — Here's How I Found Out"; Blogger, "The Network That Learns Who Comes Back".

Code and write-ups