RafinRahman/pediatric-sepsis-poc-prediction

Point-of-care ML model for predicting poor clinical outcome in paediatric sepsis in LMIC settings. No laboratory tests or imaging required. Uganda cohort. ideSHi, Bangladesh. # Predicting Poor Clinical Outcome in Paediatric Sepsis Using Point-of-Care Features **Md Rafin Rahman**¹, **Abdallah Tasawar Khan**², **Jafren Iqbal Rose**³, **Md Rofiqur Rahman**¹, **Firdausi Qadri**¹ ¹ Institute for Developing Science and Health Initiatives (ideSHi), Dhaka, Bangladesh ² Kurmitola General Hospital, Dhaka, Bangladesh ³ Shaheed Monsur Ali Medical College & Hospital, Dhaka, Bangladesh **Corresponding author:** Md Rafin Rahman, mrahman@ideshi.org --- ## Overview This repository contains the complete analysis code for the manuscript: > **Predicting Poor Clinical Outcome in Children with Suspected Sepsis in a Low-Resource Setting Using Point-of-Care Features: Model Development and Evaluation on a Ugandan Cohort** > Rahman MR, Khan AT, Rose JI. *Submitted for publication, 2026.* The study develops and evaluates a machine learning model for predicting a composite poor clinical outcome (in-hospital death or length of stay exceeding five days) in children under five years admitted with suspected sepsis in a sub-Saharan African low-resource setting. The model is constrained to point-of-care clinical features only — no laboratory tests, no imaging — collectable by a trained nurse or community health worker with a thermometer, pulse oximeter, and MUAC tape. --- ## Repository Structure ``` ├── 01_preprocess.py # Data cleaning, feature engineering, encoding, imputation ├── 02_model.py # Baseline model training and evaluation (4 algorithms) ├── 03_tune.py # Hyperparameter optimisation with Optuna (TPE sampler) ├── 04_shap.py # SHAP explainability analysis (LinearExplainer) ├── data/ │ ├── feature_list.json # Final list of 45 selected features │ └── processed.csv # Preprocessed dataset (NOT included — see Data Access below) ├── outputs/ │ ├── tuned_results.json # Final model performance metrics (holdout test set) │ ├── final_calibrated_stats.json # Platt-calibrated model stat …