P-55627590-189 - Prediction of ameloblastoma recurrence using random forest – A machine learning algorithm.

Abstract

Listen

Translate article

The purpose of this study was to investigate whether high likelihood of recurrence of ameloblastoma could be predicted using the random forest model, a machine learning algorithm. We collected data from patients treated for ameloblastoma from 1999-2019 at the University of Hong Kong. Fourteen clinical parameters were used to grow the decision trees in classifying patients with or without ameloblastoma recurrence in the follow up period. The random forest algorithm was computed 100 times in training cohort (n=100) and verified in the testing cohort (n=50). We used Receiver operating characteristic curve (ROC) and its Area under curve (AUC) as the performance measurement of separability. 150 patients (76 female and 74 male) were recruited with mean follow-up time of 103 months. A total of 25 cases recurred (16.7%) within the past 20 years. The AUC was calculated for the median and mean ROC curve which was 0.777 and 0.825 respectively. The results showed that the random forest model was able to predict recurrence of ameloblastoma with reliable accuracy. The four most important variables influencing ameloblastoma recurrence were the time elapsed from treatment, initial surgical treatment, tumor size and radiographic presentation. This provides insight in detecting high-risk patient groups to monitor recurrence. Further application of random forest to other diseases can greatly benefit clinical decision. The purpose of this study was to investigate whether high likelihood of recurrence of ameloblastoma could be predicted using the random forest model, a machine learning algorithm. We collected data from patients treated for ameloblastoma from 1999-2019 at the University of Hong Kong. Fourteen clinical parameters were used to grow the decision trees in classifying patients with or without ameloblastoma recurrence in the follow up period. The random forest algorithm was computed 100 times in training cohort (n=100) and verified in the testing cohort (n=50). We used Receiver operating characteristic curve (ROC) and its Area under curve (AUC) as the performance measurement of separability. 150 patients (76 female and 74 male) were recruited with mean follow-up time of 103 months. A total of 25 cases recurred (16.7%) within the past 20 years. The AUC was calculated for the median and mean ROC curve which was 0.777 and 0.825 respectively. The results showed that the random forest model was able to predict recurrence of ameloblastoma with reliable accuracy. The four most important variables influencing ameloblastoma recurrence were the time elapsed from treatment, initial surgical treatment, tumor size and radiographic presentation. This provides insight in detecting high-risk patient groups to monitor recurrence. Further application of random forest to other diseases can greatly benefit clinical decision.