Advancing electrochemical biosensors using machine learning

Machine learning (ML) has achieved impressive strides in many scientific domains as an advanced data processing tool. ML for image processing, natural language processing, and audio/video processing is already an industry standard. However, very little progress has been made to transform these techniques into the field of bio-sensing, especially electroanalytical sensing systems. Electrochemical sensors, especially ones designed to sense in complex media such as biofluids are inherently noisy and are prone to fouling and drift, decreasing their accuracy and preventing their prolonged use. ML has been theorized to be a helpful tool to mitigate these problems to an extent. In this work, we focus on printable, low-cost, non-enzymatic sensors to detect small molecules such as neurotransmitters in biofluids. Using functionalized laser-induced graphene as a testbench, we demonstrate that multi-modal data fusion as well as multi-peak ML-based analyte quantification improve the limit of detection (LOD) of the sensors by up to two orders of magnitude in comparison to conventional single-mode single-peak analysis. Furthermore, one of the main reasons for the lack of application of ML in electrochemical sensors is the absence of easily available/collectible big data to train the models. To solve this issue, we have designed and custom-built low-cost multichannel potentiostats capable of collecting data from more than a hundred sensors at once.
To supplement this capability, we also built automated liquid handling robots; custom built to perform high throughput electrochemical sensor benchmarking. Using this data, we demonstrate the proof-of-concept application of various ML models to improve sensor LOD and mitigate sensor drift. Collecting prolonged dopamine redox reaction data on the same sensor over days, we characterize the sensor drift and fouling in
LIG sensors. While conventional techniques fail to correct this drift sufficiently, we demonstrate the use of a neural network model to successfully use the same sensor up to 5x longer. Hence, ML serves as a powerful tool to augment and address some of the commonly known challenges with biosensors and to bridge the gap between the lab and the public for biosensor applications, toward calibration-free and long-term monitoring.