Robust and Explainable Multi-Disease Prediction Under Incomplete and Noisy Symptom Information

Symptom-based disease prediction models are usually evaluated with complete and error-free inputs, although patient reports are often incomplete or inaccurate. This study evaluated how Bernoulli Naive Bayes, Logistic Regression, Random Forest, Extreme Gradient Boosting (XGBoost), and Support Vector Machine (SVM) respond to controlled symptom corruption and whether their explanations remain stable. After exact duplicates were removed, a public benchmark of 4,920 records yielded only 304 unique symptom patterns for 41 diseases, evaluated with ten-times-repeated stratified five-fold cross-validation. Missing, false-negative, and false-positive symptoms were injected into the validation folds at 0%–40%, with five realizations per level. Robustness was summarized by the normalized area under the Macro-F1 robustness curve (nAURC), computed by trapezoidal integration divided by the corruption range, so that 1.0 indicates no loss. Random Forest was the most robust to missing and false-negative symptoms (Macro-F1 of 0.8722 and 0.8663 at 40%). False-positive symptoms were the most harmful and caused the SVM to collapse into near-constant predictions. Logistic Regression achieved the highest mean nAURC (0.8375), followed by Bernoulli Naive Bayes (0.8287). SHapley Additive exPlanations (SHAP) rankings were least stable under false-positive noise. Because the benchmark contains few, highly discriminative patterns, the findings indicate relative robustness rather than clinical diagnostic accuracy.