2159P - Machine learning algorithms to predict the risk of chemotherapy-related toxicity in older Asians

Background

The International Society for Geriatric Oncology (SIOG) recommends the use of geriatric assessments (GA) or clinical adverse event risk predictors when managing older cancer patients. However, traditional GA and risk predictors are rarely utilized in the real-world setting due to limitations in time and resources. The objective of this study was to assess the performance of machine learning algorithms in predicting the risk of post-chemotherapy grade 3-5 toxicity in older Asian patients.

Methods

This study is based on a dataset of patients from a prospective study at the National University Cancer Institute, Singapore (NCIS) that successfully validated the CARG model in older Asian patients. In this study, 200 patients aged 70 years or older of various Asian ethnicities, with a solid tumor diagnosis and undergoing chemotherapy were prospectively recruited over the period of June 1, 2017, to January 1, 2019. We trained four machine learning models – Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting (Xgboost), and Light Gradient Boosting Machine (Light DBM) to predict the outcomes of grade 3-5 (CTCAE version 4.0) adverse events post-chemotherapy.

Results

The median patient age was 74 years (IQR, 71 to 78), and 110 (55.0%) were male. 137 (68.5%) patients experienced grade 3 to 5 chemotherapy toxicity. The best-performing model for predicting grade 3-5 chemotherapy toxicity in the test set was the Xgboost model which achieved an area under the ROC curve (AUC) of 0.87. By comparison, the CARG clinical model and oncologist prediction of grade 3-5 toxicity achieved an AUC of 0.61 and 0.74 respectively in the test set. The best performing Xgboost model included elements of the CARG model in addition to clinical features such as the Karnofsky Performance Scale, number of co-morbidities and medications, current treatment line, “Timed-Up and Go”.

Conclusions

The Xgboost Machine Learning model was able to predict the occurrence of grade 3-5 toxicity post-chemotherapy in elderly Asian patients with a better performance compared to the CARG clinical model and oncologists’ predictions. These results will be further evaluated in external, prospective patient cohorts to validate them for clinical use.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

National University Cancer Institute, Singapore.

Disclosure

All authors have declared no conflicts of interest.