Application of machine learning methods for the diagnosis of Lyme disease with a fluorescent plasmonic biosensor

Lyme disease is a tick-borne illness caused by Borrelia burgdorferi that affects approximately 476,000 people every year in the United States. It is a significant clinical challenge due to the variety and inconsistency of symptoms. The currently approved diagnostic approach has significant drawbacks, such as low accuracy. There is a need for improved serological testing. A grating-coupled fluorescent plasmonic biosensor overcomes many of those drawbacks by using nanoscale plasmonic gratings to enhance signal from antibodies against B. burgdorferi antigens printed in a microarray. The resultant data have been analyzed with receiver operator characteristics (ROC) to determine diagnostic thresholds. These were compared to various machine learning methods, including support vector machines (SVM), binary logistic regression (B-LR), multiclass logistic regression (M-LR), and multilayer perceptron (MLP). Each method was evaluated on its sensitivity and specificity. The data were split into four classes, these being early Lyme, late Lyme, healthy negative, and ‘look-alike’ diseases. Overall, M-LR gave the highest accuracy, with 94%. It scored best with early Lyme and second-best with ‘look-alike’ diseases, both of which are often missed by the STTT. The M-LR technique had the highest sensitivity with 95% but underperformed both the standard two tier test and the ROC on specificity. The difference was small, but important in the context of clinical usability. The clinical applications may also be limited by the lack of transparency inherent in ML techniques. A final caveat was the small sample set used to achieve these results. More samples will likely increase the accuracy of the models. This approach represents a novel approach to Lyme diagnostics and demonstrates the potential to improve clinical outcomes.