A Machine Learning Classifier to Identify Lung Transplant Recipients (LTRs) Likely to Develop Invasive Aspergillosis (IA) One Year After Transplantation

Purpose This study investigates the risk of developing invasive aspergillosis (IA) in LTRs in the first year of follow-up. We developed and validated 3 diverse machine learning (ML) models to identify patients likely to develop IA in the first year after lung transplantation. Methods A total of 791 LTRs from January 2010 to January 2017 were followed-up for 1 year after transplantation. IA diagnosis was established as per ISHLT criteria. The data consisted of 13 variables listed in Table 1. Based on transplantation dates, we divided the cohort into 553 and 238 cases for development and validation respectively. We used 3 diverse classification methods (Naïve Bayes, Decision Tree, and Simple Logistic regression) to construct ML classification models. Results The use of statins and the presence of Aspergillus colonization post-lung transplant present strong indicators related to IA 1 year after lung transplant (Table 1). The Naïve Bayes classifier (Table 2) achieved sensitivity of 83.3% (CI95% 52-98), specificity 66.4% (CI95% 60-73) and AUC of 86.3% (CI95% 73-100) and presented the most consistent classification performance between development and validation as shown. All 3 ML methods independently utilized the same predictor variables and achieved similar prediction performance (Table 2). Conclusion The validation of 3 independent classification models showed that the Aspergillus colonization was indicative of the development of IA and use of statin was associated with a fewer cases of IA. Further clinical validation to assess the utility of using these models is warranted. This study investigates the risk of developing invasive aspergillosis (IA) in LTRs in the first year of follow-up. We developed and validated 3 diverse machine learning (ML) models to identify patients likely to develop IA in the first year after lung transplantation. A total of 791 LTRs from January 2010 to January 2017 were followed-up for 1 year after transplantation. IA diagnosis was established as per ISHLT criteria. The data consisted of 13 variables listed in Table 1. Based on transplantation dates, we divided the cohort into 553 and 238 cases for development and validation respectively. We used 3 diverse classification methods (Naïve Bayes, Decision Tree, and Simple Logistic regression) to construct ML classification models. The use of statins and the presence of Aspergillus colonization post-lung transplant present strong indicators related to IA 1 year after lung transplant (Table 1). The Naïve Bayes classifier (Table 2) achieved sensitivity of 83.3% (CI95% 52-98), specificity 66.4% (CI95% 60-73) and AUC of 86.3% (CI95% 73-100) and presented the most consistent classification performance between development and validation as shown. All 3 ML methods independently utilized the same predictor variables and achieved similar prediction performance (Table 2). The validation of 3 independent classification models showed that the Aspergillus colonization was indicative of the development of IA and use of statin was associated with a fewer cases of IA. Further clinical validation to assess the utility of using these models is warranted.