PredictMod: A Machine Learning Application for Predicting Patient Response to Interventions for Prediabetes
Open AccessThe prevalence of prediabetes in the United States has risen to 38% in 2023, with the vast majority unaware of their condition. Consistent elevation of blood glucose in conditions like prediabetes and Type 2 diabetes (T2D) often lead to more severe outcomes such as heart disease and stroke. Therefore, providing preventative care to patients diagnosed with these conditions is critical. Several studies indicate an association between gut dysbiosis and the progression of prediabetes to T2D. Additionally, it is well documented that lower beta diversity through reduction in specific microbial taxa (e.g., butyrate-producing bacteria) is associated with the onset of prediabetes. This connection provides a conduit for identifying signals within the gut microbiome of prediabetic individuals that can aid in predicting patient outcomes prior to intervention. Machine learning (ML) can be used to detect these microbial signatures and identify key patterns for predicting intervention outcome, particularly when trained with data from patients undergoing dietary modifications or implementing regular exercise. Here, we demonstrate a tool called PredictMod which had previously been developed to predict epilepsy intervention outcomes. This was accomplished by using publicly available metagenomic data obtained from NCBI’s sequence read archive (SRA) paired with prediabetic patient electronic health records (EHR) to predict disease outcomes and treatment response from baseline gut microbiome signatures. The ML-based model using EHR data was generated using synthetic patient health records to predict response status, as this data was more readily available and assisted with concept validation. The best performing algorithms were bagged tree ensembles and neural networks with regard to the predicting patient outcomes from metagenomic and EHR data, respectively. Specifically, the algorithms trained on metagenomic and synthetic patient data were able to perform at 88.9 and 83% accuracy, respectively. Furthermore, precision scores were 80% for both decision tree and neural network models, which establishes both models can predict response status with moderate confidence. These methods and models serve as an extension of, PredictMod, which serves to aid in clinician decision-making with regard to interventions for prediabetes.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.