Abstract Body (Do not enter title and authors here): Introduction: Early and accurate classification of transthyretin amyloidosis (ATTR) is crucial for improving patient outcomes. However, nonspecific symptoms and heterogeneous disease variations have made ATTR diagnosis challenging. Leveraging advancements in machine learning (ML) and large language models (LLMs), this study aims to enhance diagnostic accuracy by analyzing electronic health records (EHRs) data.
Hypothesis: Can the integration of innovative feature formulation and large language models improve the performance of ML models in diagnosing ATTR using EHR data?
Goals/Aims: Our primary goal was to develop a robust ML model that classifies the presence of ATTR in patients with high precision for clinical implementation. We aimed to utilize advanced feature formulation techniques and LLMs to capture complex patterns in patients’ EHR data.
Methods: A cohort of 1,158,426 patients from University of Pennsylvania Health System was retrospectively analyzed, of whom 261 ATTR patients and 52,200 patients sharing at least one comorbidity were selected for ML model training. Appropriate data engineering was conducted to handle missing values and feature construction. We innovated EHR feature formulation, which incorporated clinical, demographic, and patient history data. Moreover, we integrated an LLM to extract meaningful patterns and context from unstructured data. Various ML algorithms were evaluated, with hyperparameter tuning performed to optimize model performance.
Results: The developed model demonstrated a performance of F1 score 0.6, along with a precision of 74% and sensitivity 51% and an area under the precision-recall curve (AUPRC) of 0.5. These results highlight the model's potential in reliably classifying ATTR. The next step involves the clinical evaluation of patients, including recalling patients for technetium scan and ATTR genetic testing to confirm the diagnosis.
Conclusion(s): Our study demonstrates that the integration of innovative feature formulation and large language models can substantially enhance the performance of machine learning models in predicting ATTR. This approach not only enhances diagnosis rate but also provides a scalable solution for clinical implementation. Further investigation by clinicians will validate the model's effectiveness in a real-world setting, potentially transforming ATTR diagnosis and patient management.