Advancing electrochemical biosensors with AI-enhanced post-processing methods and portable electronic transducers

Electrochemical biosensors (EBs) are among the most widely researched biosensing technologies due to their simplicity, cost-effectiveness, and compatibility with a wide range of nanomaterials. Their potential as point-of-care (PoC) diagnostic tools has attracted increasing attention, enabling rapid and on-site analysis for healthcare and environmental monitoring. However, the progression of EBs from laboratory prototypes to practical applications remains limited. This challenge arises mainly from inefficient post hoc analysis methods that rely heavily on manual chemometric approaches, and the lack of affordable and versatile portable transducers that are capable of supporting diverse electroanalytical techniques. Moreover, while artificial intelligence (AI) has emerged as a promising tool for data-driven analysis, its adoption in EB research is hindered by the wide range of available algorithms and preprocessing techniques, making it difficult for non-specialists to select suitable frameworks. There is also considerable uncertainty regarding whether conventional data treatment methods should be retained when deploying AI models. Additionally, the scarcity of data, arising from the complex and costly nature of experimentation, contributes to issues such as overfitting in AI models, further complicating the model development process. To address these challenges, this thesis investigates the applications of AI, particularly machine learning (ML), to enhance EB post hoc analysis and support their transition toward practical PoC deployment. Various feature extraction methods and ML algorithms were systematically evaluated to determine optimal configurations for classification and regression tasks. The study also examined whether conventional data preprocessing methods, which are commonly employed in chemometrics, remain necessary for ML-based approaches. To overcome data scarcity and overfitting, several regularisation methods were investigated, including a bioinspired technique based on adult neurogenesis (AN), and compared against conventional approaches such as dropout, weight decay, and others. In addition, a conditional variational autoencoder (CVAE) was employed to generate synthetic EB data for data augmentation, improving model robustness and recover model performance during extremely data-scarce conditions. The developed ML framework was validated using three case studies: (1) classification of dengue virus (DENV) serotypes, (2) quantification of acetaminophen (ACE), and (3) detection of carcinoembryonic antigen (CEA). The DENV dataset was selected for serotype classification, as secondary DENV infections can be fatal. The ACE and CEA datasets were chosen as representatives of small molecules and biomarkers, respectively, for the regression task, specifically for predicting concentrations. Concentration is critical in both cases, where excessive ACE levels indicate potential overdose, while elevated CEA levels may signal the presence of disease. For classification, the developed model achieved 100% accuracy using principal component analysis (PCA) and a multilayer perceptron (MLP). For regression tasks, the combination of discrete wavelet transform (DWT) and MLP yielded the best results for ACE quantification, achieving an R² of 0.995, and the combination of PCA and MLP again achieved the best results with R² of 0.960 for CEA detection. The findings also reveal that conventional EB preprocessing is unnecessary for ML-based post hoc analysis when appropriate feature extraction methods are applied. These optimal models were further deployed in the investigation of model regularisation. When the conventional regularisation methods show insignificant improvements, the AN regularisation method further improved the models’ performance, increasing R² values to 0.999 and 0.997 for DPV and EIS datasets, respectively. Additionally, synthetic data generation enhanced model performance by reducing mean squared error by up to 63.77% and enabled comparable performance with half the original training data size. In summary, this thesis establishes a comprehensive framework for integrating AI into electrochemical biosensing, replacing the conventional chemometric analysis with modern data-driven intelligence. It provides systematic insights into the selection of suitable preprocessing, feature extraction, and ML algorithms for both classification and regression tasks, clarifying when conventional steps can be omitted without compromising model performance. The effective yet popular AN regularisation approach and the use of synthetic data generation via CVAE contribute new methodological directions for handling small EB datasets, enhancing model robustness and generalisability. Furthermore, the development of a smartphone-based, multi-technique potentiostat demonstrates the feasibility of translating laboratory-grade electrochemical analysis into a portable, cost-effective, and adaptable platform for PoC applications. Collectively, these contributions accelerate the realisation of intelligent, accessible, and scalable EB systems, paving the way toward autonomous biosensing and more effective data-driven diagnostics.