Improvement of Adequate Digoxin Dosage: An Application of Machine Learning Approach.

Area of Science:

• Pharmacology and Clinical Pharmacy
• Medical Informatics and Machine Learning

Background:

• Digoxin is a high-alert medication due to its narrow therapeutic index and potential for drug-to-drug interactions (DDIs).
• Preventable digoxin toxicity affects approximately 50% of cases, highlighting the need for improved initial dosing strategies.
• Accurate initial digoxin dosing is crucial for optimizing therapeutic outcomes and patient safety.

Purpose of the Study:

• To apply machine learning techniques for predicting the appropriateness of initial digoxin dosage.
• To evaluate the performance of various machine learning models in determining correct digoxin dosing.
• To assess the impact of drug-to-drug interactions on the accuracy of dosage prediction models.

Main Methods:

• Utilized data from 307 inpatients treated with digoxin between 2004 and 2013.
• Collected ten independent variables including demographics, laboratory data, and congestive heart failure (CHF) status.
• Employed six machine learning algorithms: decision tree (C4.5), k-nearest neighbors (kNN), CART, randomForest (RF), multilayer perceptron (MLP), and logistic regression (LGR) using Weka 3.7.3 software.

Main Results:

• Random Forest (RF) demonstrated excellent performance (AUC 0.912) in the non-DDI group, followed by MLP (0.813).
• RF also achieved the best performance (AUC 0.892) in the DDI group, with CART and MLP showing strong results.
• Decision tree-based approaches (C4.5, CART) and MLP consistently exhibited superior accuracy irrespective of DDI status.

Conclusions:

• Machine learning, particularly decision tree-based methods and MLP, can accurately predict appropriate initial digoxin dosage.
• These data mining techniques offer a valuable supplementary tool for clinicians to enhance digoxin safety.
• Developing a dosage decision support system can significantly improve clinical practice and reduce digoxin-related adverse events.