Prediction of High-risk Group for Dementia Based on Life-log

Abstract

This study proposes a framework that learns characteristics for classifying high-risk for dementia and normal subjects inherent in the life log data and improves accuracy by using only some of the data. In Chapter 1, a Bi-LSTM-based basic model structure was proposed for the classification of high-risk group for dementia. As a result of evaluating performance through k-fold cross validation, there was a big difference in performance for each split. Average performance was poor, but it performed well in some splits. Through this, we thought that the developed model could only classify data under specific conditions well. Even if it takes more time, since a large number of data can be acquired from the subject, if some of them can be diagnosed, the subject also can be diagnosed. Therefore, we propose selective diagnostic methods that increase the accuracy even if the number of diagnosable data is reduced. In Chapter 2, we propose a selective diagnostic method by model classifiable data filtering. For selective diagnosis, a framework consisting of a filtering model and a classification model is used. The filtering model classifies only data with confidence above the specified threshold as diagnosable data. The classification model classifies diagnosable data as cognitive normal or cognitive impairment. This method was able to classify 34.7% of the subjects with an accuracy of 69.7%. Chapter 3 covers selective diagnostic methods using model calibration. Because the purpose of model calibration is to make the prediction probability equal to the accuracy, selective diagnosis with high accuracy is possible by diagnosing only data with a confidence over the threshold in a well-calibrated model. Label smoothing and temperature scaling were used to calibrate the model. As a result, by calibrating the model using label smoothing, it was possible to diagnose 45.7% of the subjects in the data set with accuracy of 66% and AUC of 0.696. and by calibrating the model using temperature scaling, it was possible to diagnose 42.5% of the subjects in the data set with accuracy of 68.9% and AUC of 0.640.