Functionalized Nanofinger Enhances Pretrained Language Model Performance for Ultrafast Early Warning of Heart Attacks

Heart attacks are the leading cause of death worldwide, which means an accurate early warning system is needed. Traditional methods, such as an electrocardiogram (ECG) and blood test, usually require expert interpretation and take more than 15 min to obtain diagnostic results, which often results in delayed treatment. Our previous work [Liu, Z. <i>Small</i> 2023, 19(2), e2204719] developed a functionalized nanofinger platform utilizing Raman spectroscopy and machine learning (ML) to detect heart attack-related biomarkers, brain natriuretic peptide (BNP) in blood samples, achieving a high true positive rate (98%) but with limited true negative accuracy. This study enhances diagnostic accuracy by integrating the AI-driven analysis of patient symptoms with blood biomarker analysis. We incorporate additional patient features, including current and historical symptoms. We fused the biomarker test results, which is a one-dimensional probability vector ranging from 0 to 1, with patient symptom description as part of the sentence, which is then processed by a fine-tuned pretrained language model to generate embeddings and passed to a classification head for diagnosis. For data augmentation, we employ generative models to synthesize realistic patient cases, expanding the data set and improving model robustness. Our approach achieves an accuracy rate of 99.19%, outperforming conventional diagnostic methods. Given the difficulty of obtaining sufficient clinical data, this study presents a scalable AI-based solution for early-stage heart attack detection by integrating functionalized nanofinger-based blood analysis with large language models (LLMs) and leveraging generative models to synthesize realistic patient cases for data augmentation. This approach significantly advances the automation of diagnosis in modern healthcare and can be readily adapted to other disease detection tasks.