Machine learning applied in patient-reported outcome research-exploring symptoms in adjuvant treatment of breast cancer

BACKGROUND: Patient-reported outcome (PRO) data may help us better understand the life of breast cancer patients. We have previously collected PRO data in a national Danish breast cancer study in patients undergoing adjuvant chemotherapy. The aim of the present post-hoc explorative study is to apply Machine Learning (ML) algorithms using permutation importance to explore how specific PRO symptoms influence nonadherence to six cycles of planned adjuvant chemotherapy in breast cancer patients.

METHODS: We here investigate ePRO-data from the 347 patients. The ePRO presented 42 PROCTCAE questions on 25 symptoms. Patients completed the ePRO before each cycle of chemotherapy. Number of patients with completion of the scheduled six cycles of chemotherapy were registered. Two ML models were applied. One aimed at discovering the individual relative importance of the different questions in the dataset while the second aimed at discovering the relationships between the questions. Permutation importance was used.

RESULTS: Out of 347 patients 238 patients remained in the final dataset, 15 patients dropped out. Two symptoms: aching joints and numbness/tingling, were the most important for dropout in the final dataset, each with an importance value of about 0.04. Model's average ROC-AUC-score being 0.706. In the second model a low performance score made the results very unreliable.

CONCLUSION: In conclusion, this explorative data analysis using ML methodologies in an ePRO dataset from a population of women with breast cancer treated with adjuvant chemotherapy unravels that the symptoms aching joints and numbness/tingling could be important for drop out of planned adjuvant chemotherapy.