XAI Implementation on Preliminary Data Analysis Phase: Explainable Output Application with Prediction of Diabetes Mellitus at Early Stage

Abstract

Background: This study aims to create a machine learning model that produces explainable, interpretable, and
trustable predictions for diabetes using an XAI approach. Objective: In the study, we have utilized an earlier approach to
implementing explainable Machine Learning. Methods: In order to apply XAI technique, we follow a brief version of CRISPDM.
(i) Data Understanding, (ii) Data Preparation, (iii) Model Planning and Building (iv) SHAP Implementation for
Interpretability. Results: Global interpretability shows us that two major contributors are symptoms of Polydipsia and Polyuria.
An algorithm doesn't "know" prior information, which is highly specific domain knowledge. Local interpretability-based
single-instance explanation showed decent multivariate reasoning capability. If the reasoning was based on a simple univariate
approach, positive polyuria alone should result in a high probability of positive model output, considering the positive SHAP
value of polyuria. Conclusion: The model output results 99.7% confidence to be classified as negative makes much sense since
polyuria is also a common symptom of many different situations, such as diabetes insipidus, Kidney disease, Liver failure,
Medications that include diuretics, Chronic diarrhea, Cushing’s syndrome, Psychogenic polydipsia, Hypercalcemia,
Pregnancy.