Ablation studies towards interpretable ensemble deep neural networks for mental health classification

Post-pandemic, the prevalence of mental health issues has markedly increased within the IT sector due to the extensive implementation of work-from-home protocols. In response to this trend, various industries have instituted an array of wellness initiatives for employees, aligning these efforts with global sustainable development goals. Multiple factors impact mental health, including family history, interactions with colleagues, and current mental health status, encompassing personal, professional, and social dimensions. The exceptional flexibility and adaptability of ensemble methods and deep learning models have facilitated their widespread application in mental health research. This study presents an ensemble bagging deep learning approach employing recurrent neural networks (LSTM (Long Short-Term Memory), GRU (Gated Recurrent Unit), and BiRNN (Bidirectional RNN)) to evaluate the current mental health of IT professionals. The proposed methodology is applied to the publicly accessible open-source Mental Illness $\mathbf{2 0 2 1 + 2 0 2 2 + 2 0 2 3}$ dataset pertaining to IT industry employees. This research revealed that among the selected 37 features, the proposed ensemble deep neural network method achieved a final performance metric (accuracy) of approximately 90.13%. To tackle the interpretability challenge inherent in deep learning models, ablation studies are performed to elucidate the model’s behavior, analyzing the impact of individual hyperparameters on model predictions and exploring the interconnections between feature importance and performance metrics. The investigation describes that augmenting the values of the features, dropout layer, and epochs contributes to enhancing the model’s performance. However, variations in hidden units did not yield significant improvements.