HEMA_CLASS

Hemaclass XAI: Deep Stacking Ensemble for Malaria and Sickle Cell Anemia

Model Description

The Hemaclass XAI model is a phase-4 prototype Clinical Decision Support System (CDSS) designed to classify Malaria, Sickle Cell Anemia (SCA), Co-infections, and Negative (Healthy/Other) patient states. Developed specifically for deployment in resource-constrained settings in Western Kenya, the model utilizes a robust Deep Stacking Ensemble architecture coupled with SHAP (SHapley Additive exPlanations) for transparent, clinician-friendly interpretability.

• Developer: Mabonga Labs / Hemaclass Project
• Model Type: Multi-Class Tabular Classification (Deep Stacking Ensemble)
• Primary Architecture: Random Forest, Support Vector Machine (RBF), and XGBoost (Base Learners) -> Logistic Regression (Meta-Learner).
• Hyperparameter Tuning: Nested Cross-Validation with Bayesian Search (skopt).

Intended Use & Target Audience

• Primary Use Case: Triage and secondary diagnostic validation for clinicians operating in malaria-endemic regions with high SCA prevalence.
• Target Audience: Medical doctors, clinical officers, and healthcare technicians.
• Out-of-Scope Uses: This model is not a standalone diagnostic device. It is designed for decision support. It should not override human clinical judgment or be used as a replacement for definitive laboratory protocols (e.g., blood smears, Hb electrophoresis).

Clinical Protocol & Hardcoded Overrides

To prioritize patient safety, the inference pipeline integrates hardcoded clinical rules that override the AI's probability outputs in critical, life-threatening scenarios:

1. Severe Hyperhemolytic Crisis: Hemoglobin (Hb) < 5.0 g/dL.
2. Acute Hemolytic Malarial Crisis: Reticulocyte Count > 8.0% + Positive Malaria RDT.
3. Rapidly Progressing Vaso-occlusive Malarial Crisis: Rapid Hb decline (>1.5g/dL in 48h) + Positive Malaria RDT + Presence of HbS genotype. When triggered, the system flags the diagnosis as "Co-infection" with 100% confidence and alerts the clinician to admit the patient to a high-dependency unit.

Model Input Features

The model ingests 24 clinical biomarkers and autonomously engineers 3 derived features:

• Demographics: Age, Sex
• Vitals & Symptoms: Body Temperature, Fever, Chills, Headache, Muscle Aches, Fatigue, Loss of Appetite, Jaundice, Abdominal Pain, Joint Pain, Splenomegaly, Severe Pallor, Lymphadenopathy.
• Laboratory Markers: Malaria RDT (Binary), Hemoglobin (Hb), WBC Count, Platelet Count, Reticulocyte Count, Rapid Hb Decline Alert.
• Hemoglobin Fractions: HbA, HbS, HbF.
• Engineered Features: Symptom Severity Score, Age Group (Categorical), Infection-to-Anemia Ratio (WBC / Hb).

Data Preprocessing & Augmentation

• Missing Data: Handled using Multiple Imputation by Chained Equations (MICE) up to 30 iterations to preserve complex clinical covariances.
• Class Imbalance: The foundational dataset (~350 retrospective patient records from Western Kenya) was highly imbalanced. Extreme SMOTE (Synthetic Minority Over-sampling Technique) was applied to map boundaries and synthesize a robust, balanced training matrix of 6,000 clinical profiles (1,500 per target class).
• Encoding: Deterministic Ordinal Encoding for categorical values; Z-Score Normalization for numerical features.

Explainability (XAI)

The system integrates SHAP (TreeExplainer) applied to the XGBoost component of the stacking ensemble. Local explanations (Waterfall plots) are generated for every inference, providing clinicians with exact quantification of how individual biomarkers (e.g., HbS% or Symptom Severity) contributed to the final predicted risk score.

Evaluation & Metrics

The model was evaluated on an isolated, unseen clinical test set prior to SMOTE augmentation.

• Metrics Tracked: Macro-F1 Score, Macro-Sensitivity (Recall), Macro-Specificity, and overall Accuracy.
• Statistical Significance: Friedman/Nemenyi post-hoc testing confirmed the Stacking Ensemble significantly outperforms isolated baseline models (p < 0.05). (Note: Refer to the specific model logs or the connected Gradio dashboard for live metrics).

Ethical Considerations & Limitations

• Geographic Bias: The base dataset reflects the epidemiology of Western Kenya. Prevalence features (like overlapping splenomegaly in SCA and Malaria) may not generalize accurately to populations outside of Sub-Saharan Africa or regions with varying Plasmodium falciparum endemicity.
• Synthetic Data Artifacts: Because SMOTE was used extensively to augment the data for deep learning convergence, extreme edge-case boundaries may occasionally display synthetic bias. Prospective validation on a large-scale real-world clinical trial is required before Phase 5 medical device certification.
• Explainability Proxy: The SHAP explanations are derived from the XGBoost sub-estimator rather than the entire Stacking Classifier boundary, serving as a highly accurate proxy rather than an absolute mathematical reflection of the meta-learner.

How to Get Started

To interact with the model via the UI, please visit the connected Hugging Face Space.

For programmatic inference using Python:

import joblib
import pandas as pd

# 1. Load Artifacts
model = joblib.load('ensemble_model.pkl')
scaler = joblib.load('scaler.pkl')
imputer = joblib.load('imputer.pkl')

# 2. Prepare patient data (ensure columns match FEATURE_NAMES)
# patient_df = pd.DataFrame({...})

# 3. Preprocess
# X_imp = imputer.transform(patient_df)
# X_scaled = scaler.transform(X_imp)

# 4. Predict
# predictions = model.predict(X_scaled)
# probabilities = model.predict_proba(X_scaled)