Comparison of Machine Learning Methods in Mild Cognitive Impairment Prediction for Cancer Patients Using EHR Data

Abstract

AbstractCancer and dementia are common in aging populations. Mild cognitive impairment (MCI) is a stage between the cognitive changes of normal aging and dementia that can lead to a decline in quality of life. With the substantial improvement of survival in many cancers, maintaining a high quality of life has become a new goal in cancer care. Identifying those patients with a high risk of developing MCI may facilitate early intervention and further improve patient care. The objective of this study is to survey machine learning techniques and AutoML to model the early detection of MCI in patients with cancers using the features which are known risk factors in dementia and accessible in the electronic health records (EHR). We compared multiple machine learning methods and explored AutoML to predict 1-year risk of MCI for cancer patients. Among 27 models, XGBoost in AutoML gave the highest AUC (0.79), suggesting the superiority of using automated machine learning tools to search for the best model and parameters. The feature importance analysis revealed that cancer patients with brain malignancy, hypertension, or cardiovascular diseases are more likely to develop MCI. The overall poor performance indicates more efforts should be made to improve data quality and increase features and sample size.