Exploring adult neurogenesis for enhanced neural networks: Model regularisation approaches for electrochemical biosensor data

Recent advancements in artificial intelligence (AI) and biomedical engineering have greatly contributed to the development of machine learning (ML)-enhanced electrochemical biosensors (EBs). While the combination of ML, particularly neural networks (NNs), with EBs shows great promise, the development of ML models for post hoc analysis of EB data often faces challenges such as overfitting, mainly due to the limited dataset sizes commonly encountered in the early stages of research, mainly due to the labour-intensive and resource-heavy nature of EB development. Several recent AI-biosensor studies have included the utilisation of model regularisation techniques, such as dropout and weight decay, to improve model performance. However, there is a lack of in-depth investigation comparing the effectiveness of different regularisation methods. To address this gap, this study systematically compares various conventional regularisation methods for developing ML models for post hoc EB analysis, while also exploring an effective yet underutilised technique called Adult Neurogenesis (AN), which involves reinitializing weights of specific nodes to enhance NN generalisation. Two common EB datasets, differential pulse voltammetry (DPV) and electrochemical impedance spectroscopy (EIS), were used for this analysis. The results show that AN outperforms the conventional methods, with significant reductions in error of 68.84% and 70.35% for the DPV and EIS datasets, respectively, compared to the baseline models. Furthermore, AN achieved R² values of 0.999 and 0.997 for the DPV and EIS datasets, respectively. These findings highlight the potential of AN to substantially improve the performance of EBs, where the same approach can possibly be applied across other areas.