AI (Artificial Intelligence)-Coupled Self-Calibrating SERS Spectroscopy for Robust Clinical Diagnosis of Diabetes and Associated Complications

Diabetes mellitus and the complications associated with it have come across as the major cause of increased mortality, disability, and comorbidity across the world. Increasing burden of the diabetic patients in public healthcare systems is alarming, and it could have serious repercussions on the overall public health and welfare in coming days. Especially in standard practice, clinical diagnosis of diabetes mellitus requires a large volume of blood samples, and the detection procedure is very complex yet less efficient. In this regard, combining the power of artificial intelligence (AI) with a self-calibrating Surface-Enhanced Raman Spectroscopy (SERS) technique can provide an alternative diagnostic strategy for the robust clinical diagnosis of diabetes and its associated complications. Gold nanoparticles (AuNPs) and 4-mercaptophenylboronic acid functionalized AuNPs (AuNPs@4MPBA) were utilized for the non-specific and specific detection of diabetes in various serum samples (collected from different patients) using SERS technique. The concatenation of non-specific and specific SERS analysis was beneficial for the efficient, highly accurate diagnosis of diabetes using only a small portion of the samples. In the present study, a ResNet-LSTM multi-head-self-attention neural network was judiciously integrated with the self-calibrating SERS spectroscopic technique not only to identify and classify the diabetes but also to comprehend the early complications associated with it. The evaluation model was applied to the sample datasets comprising both the pre- and post-medication data of the patients obtained from a hospital. The diagnostic method was also instrumental in effectively classifying various types of diabetes (type 2, type 4, and type 7). Specifically, the concatenated SERS spectral data were classified with 98.5%, 94.6%, and 98.1% accuracy, respectively. Though the conventional diagnostic methods failed to accurately diagnose post-medication complications, our newly designed self-calibrating diagnostic model was capable of consistently enabling precise diagnosis of diabetes, particularly when the patients had a history of pre-medication. Furthermore, the self-calibration aptitude of this clinical diagnostic approach using the Cosine’s similarity and Pearson’s correlation methods provided an excellent scope for generalizing the detection method to achieve more accurate diagnostic information and to facilitate correction in cases of clinical misdiagnosis.