Leveraging Transfer Learning for Screening of Geriatric Depression

Depression among the geriatric population is a prevalent and critical public health concern that often goes undetected, and untreated, leading to adverse effects on their quality of life. Transfer learning, traditionally prominent in image and text data, is extended to tabular data to leverage knowledge gained from one task to improve predictions on a related but distinct task. The study utilizes diverse datasets encompassing socio-demographic, health-related, and psychosocial attributes of geriatric individuals, reflecting the complexity of depression prediction. The research workflow involves a three-phase process: pretraining, transfer learning, and comparative evaluation with traditional machine learning models. In the pretraining phase, a model is trained on a source task related to geriatric depression. This pre-trained model is then used as input for the target task: depression prediction. The fine-tuning phase adapts the source-knowledge-enhanced model to the target depression prediction task by updating its parameters on the target dataset. The performance of the transfer learning approach is compared with traditional machine learning models. The experiment showed that the Extreme Gradient Boosting algorithm trained on a larger benchmark dataset when applied as transfer learning approach for a similar task, i.e., for screening of depression among the geriatric population significantly outperforms (p-value less than 0.05) all other traditional algorithms in all performance metrics (Accuracy-0.97, Precision-0.95, Recall-0.98) compared to nearest algorithm, i.e., Random Forest (Accuracy-0.85, Precision-0.82, Recall-0.90). By harnessing the knowledge encoded in a source task and adapting it to a target task, the research has shown promising advancements in capturing complex patterns within tabular datasets, which may be particularly challenging to discern through conventional modeling approaches. The findings of this study underscore the potential of transfer learning as a valuable tool in enhancing the accuracy and robustness of predictive models for geriatric depression.