Adherence predictor variables in AIDS patients: An empirical study using the data mining-based RFM model

Abstract

Background: Highly active antiretroviral therapy (ART) is still the only effective method to stop the disease progression in acquired immunodeficiency syndrome (AIDS) patients. However, poor adherence to the therapy makes it ineffective. In this work, we construct an adherence prediction model of AIDS patients using the classical recency, frequency and monetary value (RFM) model in the data mining-based customer relationship management model to obtain adherence predictor variables.Methods: We cleaned 257305 diagnostic data elements of AIDS outpatients in Shanghai from August 2009 to December 2019 to obtain 16440 elements. We tested the RFM and RFm (R: recent consultation month, F: consultation frequency, M/m: total/average medical costs per visit) models, three clustering methods (K-means, Kohonen and two-step clustering) and four decision algorithms (C5.0, the classification and regression tree, Chi-square Automatic Interaction Detector and Quick, Unbiased, Efficient, Statistical Tree) to select the optimal combination. The optimal model and clustering analysis were used to divide the patients into two groups (good and poor adherence), then the optimal decision algorithm was used to construct the prediction model of adherence and obtain its predictor variables.Results: The results revealed that the RFm model, K-means clustering analysis and C5.0 algorithm were optimal. After three rounds of k-means clustering analysis, the optimal RFm clustering model quality was 0.8, 10614 elements were obtained, including 9803 and 811 from patients with good or poor adherence, respectively, and five types of patients were identified. The prediction model had an accuracy of 100% with the recent consultation month as an important adherence predictor variable.Conclusions: This work presented a prediction model for medication adherence in AIDS patients at the designated AIDS center in Shanghai, using the RFm model and the k-means and C5.0 algorithms.