Novel Risk Scores for Adverse Effects During Initiation of Anti-Arrhythmic Drugs: A Machine Learning-Based Statistical Approach for Remote Patient Monitoring
Restricted (Penn State Only)
- Author:
- Pallod, Aditi
- Area of Honors:
- Biotechnology
- Degree:
- Bachelor of Science
- Document Type:
- Thesis
- Thesis Supervisors:
- Jean-Paul Armache, Thesis Supervisor
Teh-Hui Kao, Thesis Honors Advisor - Keywords:
- Anti-Arrhythmic Drugs
Remote Patient Monitoring
Cardiology
Machine Learning
Regression Modeling - Abstract:
- Anti-arrhythmic drugs (AADs) are used to treat various types of cardiac arrhythmias, specifically atrial fibrillation. The initial in-hospital drug loading of anti-arrhythmic drugs is critical for effective treatment. This loading period ranges from three to seven days and is costly and burdensome for both hospitals and patients; hence, remote monitoring during initiation has become increasingly appealing. However, clinicians are concerned about the safety of outpatient monitoring due to the risks of adverse events during loading, which are well-documented. With the development of a prediction model for potential risk during monitoring, it is possible to identify patients who are at greater risk of experiencing hazardous adverse effects based on their characteristics or baseline laboratory tests. This work presents two risk score equations for adverse effects during the loading of two Class III AADs, dofetilide and sotalol. The risk scores use patient characteristics such as age, gender, dose, history of heart failure, concurrent use of calcium channel blockers, and history of hypertension to predict if a patient is at a high risk of adverse events (e.g., torsades de pointes, ventricular arrhythmia, QTc prolongation). Using these equations, patients can be sorted into a remote monitoring initiation protocol or standard hospital care based on their eligibility. This system not only results in a significant cost reduction as fewer patients are in the hospitals, but also ensures that the highest levels of safety and patient care are maintained by selecting only patients with the significantly low chances of adverse events. The risk scores were developed with machine learning techniques using two datasets of eighty-three and twenty-nine patients for dofetilide and sotalol, respectively. After these datasets were preprocessed, a mixed effects regression model was employed. The datasets were divided into a 95:5 ratio for training and validation. The models had a high accuracy rate for the risk score equations, but the designations based on the coefficients of the patient characteristics did not consistently correspond to identified correlations in a systematic review of pooled study data, FDA guidelines, and additional literature. This may be due to the small sample size of the raw data used, skewing of the parameters as a result of the significant proportions of missing data that had to be imputed, or clinical extremes present in the data as a result of extracting from case reports, among other factors. Despite the inaccuracies in the risk scores, this study presents a useful approach for creating a prediction tool for adverse events based on individual patient data using machine learning with a regression model. This work also presents a systematic review of pooled data from multiple studies for each drug, specifically showcasing the statistically significant and non-significant associations between parameters and adverse events. This work culminates with a literature review regarding the feasibility of remote patient monitoring during the initiation of dofetilide and sotalol. While the review had a small number of studies, it suggests that outpatient monitoring is possible with careful patient selection and appropriate monitoring using cardiac devices. Therefore, clinicians can use a risk prediction score to filter patients to maintain selectivity of the patient group undergoing remote monitoring, ultimately reducing patient and hospital costs substantially while ensuring that patients are treated safely.