Explainable-Acute-Leukemia-Mortality-Predictor

Explainable Acute Leukemia Mortality Predictor – Model Repository

This repository contains the trained machine learning model artifacts generated by the Explainable Acute Leukemia Mortality Predictor Hugging Face Space.

It serves exclusively as a persistent storage and versioning registry for models developed for:

Mortality risk prediction in patients with acute leukemia using structured clinical data.

This repository does not provide training or an interactive interface.

---

Relationship to the Application

Model development, validation, and prediction occur in the companion Space:

Synav/Explainable-Acute-Leukemia-Mortality-Predictor

Because Hugging Face Spaces use temporary storage, trained models are automatically:

1. Saved
2. Versioned
3. Uploaded here
4. Preserved as permanent releases

This ensures:

• reproducibility
• auditability
• long-term persistence
• external validation capability

---

Model Description

Each stored model is:

• Task: Binary mortality prediction (Yes/No)
• Algorithm: Logistic Regression (scikit-learn)
• Output: Probability of mortality (0–1)
• Explainability: SHAP feature attribution

Embedded preprocessing

Numeric variables

• median imputation
• standard scaling

Categorical variables

• most-frequent imputation
• one-hot encoding

All preprocessing steps are embedded within the pipeline to guarantee:

• identical inference behavior
• schema consistency
• zero manual preprocessing

---

Files Included per Release

Each version folder contains:

model.joblib

Complete scikit-learn pipeline including preprocessing, feature encoding, and the trained classifier. Ready for immediate inference.

meta.json

Structured metadata including:

• feature schema
• variable types
• evaluation metrics
• ROC/PR curve data
• calibration statistics
• confusion matrix
• decision curve analysis
• validation configuration

These artifacts enable full reproducibility and downstream analysis.

---

Evaluation Metrics Captured

Models are evaluated on held-out test data using clinical-grade performance criteria.

Discrimination

• ROC AUC
• ROC curve
• Precision–Recall curve
• Average Precision

Classification

• Sensitivity (Recall)
• Specificity
• Precision
• F1 score
• Accuracy
• Balanced accuracy
• Confusion matrix

Calibration

• Calibration (reliability) curve
• Brier score

Clinical Utility

• Decision Curve Analysis (net benefit)

---

Repository Structure

releases/
  └── <version>/
      ├── model.joblib
      └── meta.json

latest/
  ├── model.joblib
  └── meta.json

README.md

• releases/<version>/ → immutable historical snapshots
• latest/ → most recent validated model

---

Intended Use

These artifacts are intended for:

• Clinical research
• Risk stratification studies
• Independent external validation
• Multi-center reproducibility testing
• Educational and exploratory analysis

---

Not Intended For

These models:

• are not regulatory-approved medical devices
• do not replace clinician judgment
• should not be used for autonomous decision-making
• require local validation prior to clinical deployment

Clinical oversight is mandatory.

---

Loading a Model

import joblib

model = joblib.load("model.joblib")
proba = model.predict_proba(X)[:, 1]

No additional preprocessing is required.

---

Author

Dr. Syed Naveed Hematology & Oncology Sheikh Shakhbout Medical City Abu Dhabi, UAE

---

License

Apache 2.0

---