Deep Learning-based Prediction of Cognitive Function Using Basic Blood Test Data and NIRS-measured Cerebral Hemodynamics

Abstract

Recently, we demonstrated that deep learning allows the prediction of cognitive function using basic blood test data. In this study, we evaluated the accuracy of deep learning-based predictions of cognitive function by comparing basic blood test data and cerebral hemodynamics as measured by time-resolved near-infrared spectroscopy (TNIRS) as input data for the model. First, we used a linear regression model, random forest, and a deep neural network as contemporary machine learning regression models. We studied 202 participants to assess cognitive function using the Mini-Mental State Examination and analyzed TNIRS-measured cerebral hemodynamics, including absolute concentrations of hemoglobin, regional oxygen saturation, and optical pathlength in the bilateral prefrontal cortices at rest. The results suggested that prediction using both TNIRS and blood data inputs exhibited lower mean absolute and mean absolute percentage errors. We also confirmed that the blood test data are often useful; however, a sufficient combination, including blood counts, electrolytes, and nutrition, is required for clinical use.