Prepulse Inhibition Response Classification Model using Machine Learning Techniques

Abstract

The use of acoustic has gained attention because of its application in signal processing to analyse sound in the field of speech recognition and computational acoustic. While Machine learning (ML) techniques for classification is rapidly developing with compelling results and significant future promise, little information is known on its application in classifying startle reflex waveforms. Hence, this study was aimed at developing an improved prepulse inhibition classification model using machine learning technique. A dataset of eight hundred (800) prepulse inhibition responses were collected from the prepulse inhibition experiment performed by introducing sound at pulse alone trial 45dB, prepulse + pulse trial 85dB and randomised trials using ASLAS device. For training, validation, and testing, the data was divided into ratios of 70:15:15. Decision trees (DT) (Entropy) and (Gini), logistic regression (LR), and random forests (RF) models were selected to categorize waveforms into startle and non-startle at 85dB stimulus intensity. Performance measures were accuracy, precision, recall, and F1-score. Improved acoustic device (ASLAS) utilised RF, LR, DT (Gini) and DT (Entropy) on its dataset to yield accuracy of 96.0, 97.0, 93.0 and 94.0%; precision of 74.0, 78.0, 92.0 and 75.0%; recall of 80.0, 82.0, 75.0 and 84.0%; and f1-score, 72.0, 82.0, 83.0 and 93.9%, respectively. LR gives best accuracy (97%), DT (Gini) has highest precision (92%), while DT (Entropy) has highest Recall (84%) and F1-Score (93%). This result implies that the model predictions are correct and a high F1 score means the model performs well on both metrics precision and recall. The importance of this is that it enhanced decision-making through the elimination of blind spots by forecasting future outcomes and also helps in proactive risk management by identifying potential vulnerabilities and anomalies which minimizes losses.