A HYBRID CNN-LSTM FRAMEWORK FOR SOFTWARE DEFECT PREDICTION IN SAFETY-CRITICAL HEALTHCARE SYSTEMS WITH EXPLAINABILITY AND REGULATORY COMPLIANCE

Abstract

Lately, deeper networks - including Convolutional Neural Networks - show stronger results by uncovering layered traits across defect flows. Patterns that shift over time get captured better using Long Short-Term Memory units inside prediction setups.The performance can even become better by utilizing reverse sequences in these techniques such as BidirectionalLSTM. Combining all categories in CNN-LSTM techniques makes it possible for a system to uncover relationship which can’t be done through other tools. Usually, these models using these time patterns generally achieve higher performance over old models. It detects real cases, scoring and avoids false cases correctly. Based on the above information, this research used a combination of CNN LSTM using NASA-PROMISE datasets and patient records. After dealing with the imbalanced class issue by making modifications in these techniques, SMOTE made the performance reaches accuracy and problem detection rate of 0.94 with average deviation of 0.12. Apart from numerical assessment, the system provides trace logs according to IEC 62304 standard to implement research ideas to practical devices. Besides broad investigation on this topic, neural network approaches were applied.