Analysis of Risk Factors for Ganciclovir-Induced Thrombocytopenia and Construction of Risk-Prediction Models Using a Decision Tree Analysis
- VernacularTitle:ガンシクロビル誘発性血小板減少症の要因分析とDecision tree 解析を用いたリスク推定モデルの構築
- Author:
Shungo IMAI
1
;
Takehiro YAMADA
1
;
Kumiko KASASHI
1
;
Masaki KOBAYASHI
1
;
Ken ISEKI
1
Author Information
- Keywords: ganciclovir; data mining; decision tree analysis; decision tree model; thrombocytopenia
- From:Japanese Journal of Drug Informatics 2019;21(1):9-19
- CountryJapan
- Language:Japanese
- Abstract: Objective: Hematological toxicity, including neutropenia and thrombocytopenia, is a typical side effect of ganciclovir (GCV). We previously developed a risk-prediction model for GCV-induced neutropenia using decision tree (DT) analysis. By employing the DT model, which is a flowchart-like framework, users can predict the combination of factors that may increase neutropenia risk. However, a risk-prediction model for thrombocytopenia has not been established. Here, we aimed to identify the risk factors associated with GCV-induced thrombocytopenia and construct risk-prediction models.Method: We retrospectively evaluated the medical records of 386 patients who received GCV between April 2008 and March 2018 at Hokkaido University Hospital. Thrombocytopenia is defined as a decrease in the platelet count (PLT) to ＜50,000 cells/mm3 and to a ＜75% decrease. Risk factors of thrombocytopenia were extracted from the medical records using a multiple logistic regression analysis. Moreover, we employed chi-squared automatic interaction detection (CHAID) and classification and regression tree (CRT) algorithms to develop the DT models. The accuracies of the established models were evaluated to assess their reliability.Results: Thrombocytopenia occurred in 47 (12.2%) patients. In the multiple logistic regression analysis, data of patients with white blood cells ＜7,000 cells/mm3,PLT＜101,000 cells/mm3 and total bilirubin ≥ 0.8 mg/dL were extracted. Two risk-prediction models were constructed, and patients were divided into six and seven subgroups. In both algorithms, data on hematopoietic stem cell transplantations, PLT ＜101,000 cells/mm3, serum albumin ＜ 2.8 g/dL, total bilirubin ≥ 0.8 mg/dL, and residence in intensive care unit were extracted. The predictive accuracy of both the CHAID algorithm and the logistic regression models was 87.8% and that of the CRT algorithm was 88.3%, indicating they were reliable.Conclusion: We successfully identified the factors associated with GCV-induced thrombocytopenia and constructed useful flowchartlike risk-prediction models.
