Constructs and verifies a recency-frequency (RF) model for predicting the loss to follow-up (LTFU) in acquired immunodeficiency syndrome (AIDS) patients in China. We exported the full data of AIDS outpatients in the research unit from August 2009 to September 2020 as the observation dataset, from September to December 2020 as the prediction dataset. The data cleaned to obtain one data element per person. The classic recency-frequency-monetary (RFM) model was expanded into RFM, RF, RFL and RFML models (L: length of treatment cycle), using the k-means for model construction and C5.0 for verification to the best predictive model. The best predictive model was subjected to two rounds of k-means clustering. In the retained data were randomly divided into a training set (70%) and a testing set (30%). The artificial neural network algorithm was used to predict the LTFU of AIDS patients and the confusion matrix was used to evaluate the performance. The observation dataset included 16,949 elements and the prediction dataset had 10,748 elements. In the best predictive model, an RF model with a quality of 0.82. A retained of 13,799 data elements were retrained and randomly binned into a test set and a verification set, the accuracy rate were100.0%. LTFU of AIDS patients was predicted in the prediction set with a correct rate of 99.89% and the accuracy and precision were 85.41% and 99.76%, respectively. The RF model verified using a neural network algorithm is effective, reliable and feasible to predict the LTFU of AIDS patients in China.

Objective: AIDS patients need life-long ART, the cost of which has become a heavy medical burden on health systems. In this study, the CD4 cell count variable is modeled to subdivide the ART-treated patients, aiming to reduce the medical burden. Methods: The data of outpatients at the research unit between August 2009 and December 2020 were exported. A recency-frequency (RF) model was established to distinguish and preserve ART-treated patients. The common factor analysis (CFA) was conducted on the three indicators of the baseline/mean/last CD4 cell counts to obtain critical variables, which were then subjected to k-means modeling to subdivide ART-treated patients, and their medical burden was analyzed. Results: A total of 12,106 pieces of ART-treated patients were preserved by RF modeling. The baseline/mean/last CD4 cell counts served as important variables employed for modeling. The patients were divided into 15 types, including two types with poor compliance and poor immune reconstitution, two types with good compliance but poor immune reconstitution, four types with poor compliance but good immune reconstitution and seven types with good compliance and good immune reconstitution. The frequency of visits was 5.25-9.95 visits/person/year, and the percentage of examination fees was 44.24-59.05%, with a medical burden of 4,114-12,677 yuan/person/year, of which 42.62-70.09% was reducible. Poor compliance and poor immune reconstitution lead to excessive visits and frequent examination, which were the leading causes of the heavy medical burden of ART. Conclusion: RF modeling can be used to distinguish ART-treated patients, which can be subdivided by the CD4 cell count.

Objective To explore if CD4 can be used as an index to guide the ART compliance and predict the feasibility of post-ART immune reconstitution. Methods The data of outpatients with AIDS visiting the research unit from August 2009 to April 2021 were used. The patients were divided according to the grouping variable (CD4 in the last 1-6 consecutive times ≥ 500 cells/μL) into good immune reconstitution and poor immune reconstitution groups. Based on the baseline CD4 value, the patients were classified into 6 types. The optimal grouping variable was used to establish the post-ART immune reconstitution prediction models, and inference rules were generated. Results A total of 7,872 pieces of valid data were obtained, including 4,834 in the incomplete immune reconstitution group (CD4 ≥ 500 cells/μL or < 500 cells/μL). Taking CD4 in the last 6 consecutive times ≥ 500 cells/μL as the optimal grouping variable, 6 immune reconstitution prediction models were established, with accuracies ranging within 89.4-93.29%, and 29 inference rules were generated. Good immune reconstitution rules were more complex than poor immune reconstitution rules and had more additional conditions; the confidence was high. It was found to usually take 42 months of adherence to ART treatment to enter a good immune reconstitution state. Conclusion The model established based on the optimal grouping variable of CD4 in the last 6 consecutive times ≥ 500 cells/μL and the corresponding inference rules can effectively guide ART, assess the post-ART immune reconstitution status and lower the medical burden of ART.