Assessment of the usefulness of statistical learning systems in drawing conclusions about the cognitive performance status of elderly patients

Introduction
Quick insight into the patient’s cognitive performance can measurably help in diagnostic support and planning further care for elderly patients. Due to the ever-increasing diagnostic requirements, it can be useful to use statistical machine learning systems related to the field of machine learning. The algorithm learns the relationship between input data (test score) and corresponding output variables (diagnosis). The aim of the study was to estimate how a given learning model would cope with the estimation of the cognitive performance of an elderly patient on the basis of data collected in medical records. The aim of the study was also to compare the models used.

Material and methods
A retrospective review of the studies of 280 patients hospitalized at the Geriatric Hospital of John Paul II in Katowice in 2015-2019. Input data included biochemical indices, functional scale scores (ADL, IADL), comorbidities and sociodemographic variables. The total data was divided as follows: 2/3 study data, 1/3 test data. The following models were used: linear regression model, support vector machine, naive Bayes classifier, k-nearest neighbors method.

Results
None of the models achieved satisfactory accuracy. The best results were obtained from the linear regression model with uniform data division (AUC = 0.57, ACC = 0.60) and the linear regression model with all data learning (AUC = 0.68, ACC = 0.67).

Conclusions
Based on the collected parameters, a system of automatic inference about the patient’s cognitive performance should not be built. The linear regression model requires further empirical verifications involving larger and more diverse groups on larger and more diverse groups.