Machine Learning Algorithms in Predicting Multi-Diseases

Abstract

Most current clinical decision support with health care and machine learning are developed for a single disease and are therefore not integrated into the workflows when the patient needs to be assessed for multiple diseases. We present a single prediction architecture aimed at solving diabetes, heart disease and Parkinson's in a single integrated pipeline to overcome this limitation. However, less than 1,300 original records are provided by the three publicly available UCI benchmark sources, which is not enough to allow stable multi-model training, so we created a 50,000 -sample research dataset using merely statistical augmentation with a distribution-aware approach inside the training fold, without using real patients; this is not a 50,000 sample set of distinct real patients, it was only created for the purpose of controlled benchmarking. All the performance statistics mentioned in this document are taken from the original benchmark records that were held out, and not modified during the augmentation process. Four of these classifiers (Support Vector Machine (SVM), Decision Tree (DT), Artificial Neural Network (ANN) and Random Forest (RF)) were trained and evaluated using a common and strictly controlled experimental setup. For the three tasks, RF outperformed all the other algorithms with accuracies of $\mathbf{9 1. 9 \%,\ 8 9. 3 \%}$, and $\mathbf{9 6. 2 \%}$, and with a mean AUC-ROC of 0.955. Repeated stratified splits and paired significance tests ($p<0.05$) show that these numbers are constant between runs. One path that can be taken toward deployment is demonstrated with a Flask based web prototype. The numbers listed here are representative of a research prototype only and should not be used as evidence of clinical readiness.