1*Research Scholar, IIS (Deemed to be University), Jaipur, India. Email: sinhal.suchita@gmail.com
2Head & Associate Professor (CS & IT), Department of Computer Science & IT, IIS (Deemed to be University), Jaipur, India. Email: ruchi.nanda@iisuniv.ac.in
Background: Major Depressive Disorder (MDD) is a serious mental health condition that negatively affects a person's thoughts, feelings, and behaviors. It often leads to a persistent state of low mood, emotional disassociation, and a loss of interest in activities that were previously enjoyed. It requires accurate and timely prognosis. The increasing use of wearable devices helps in the continuous collection of actigraphy data which is a strong indicator of a depressed individual. However, the complex, non-linear, and irregular nature of actigraphy time-series makes it poorly suited for traditional machine learning techniques as well as conventional statistical approaches such as ARIMA-based models, which rely on simplified assumptions and limited temporal representation capabilities.
Methodology: To address these challenges, this study proposes DAB-HARM-Net, a hybrid ensemble framework for actigraphy-based depression prognosis. The model transcends the "single perspective" constraint of standard RNN (Recurrent Neural Network) by integrating parallel processing tiers of GRU (Gated Recurrent Unit), LSTM (Long Short-Term Memory), and BiLSTM with an Attention Mechanism and Residual Learning.
Results: Experimental results demonstrate that the proposed model outperforms traditional machine learning and standalone deep learning approaches, achieving improved robustness and generalization in handling noisy, imbalanced, and high-dimensional time-series data. This work highlights the limitations of conventional modeling techniques for actigraphy analysis and establishes the effectiveness of hybrid ensemble strategies for reliable, data-driven MDD prognosis. Evaluated on the "Depresjon" dataset, the proposed model outperforms traditional machine learning and standalone deep learning approaches. The results show that DAB-HARM-Net achieved superior predictive accuracy with a Root Mean Squared Error (RMSE) of 187.75 and a Mean Absolute Error (MAE) of 76.72.
Conclusion: These results demonstrate the model's ability to preserve clinically significant activity "extremes" and rhythmic disruptions, providing a robust, objective early-warning system for depressive episodes.
Keywords: Major Depressive Disorder, DAB-HARM-Net, LSTM, BiLSTM, GRU, Deep Learning, Digital Biomarkers, Actigraphy Data
How to cite this article: Sinhal S, Nanda R. DAB-HARM-Net: Deep Actigraphy Based Hybrid Attention Residual Metamodel for Depression Prediction. Int J Drug Deliv Technol. 2026;16(13s): 189-199. DOI: 10.25258/ijddt.16.13s.20
Source of support: Nil.
Conflict of interest: None