Predicting patient-reported symptom clusters in lung cancer patients: a machine learning approach

Abstract

Purpose or Objective Lung cancer is one of the most common cancer types in the world, with patients suffering from multiple co-occurring symptoms: i.e. ‘symptom clusters (SC)’. Identifying SC is important to anticipate on other symptoms within a cluster and to uncover possibly overlooked symptoms. Also, supportive care interventions should aim to target multiple symptoms within a SC by addressing 1 or 2 symptoms and therefore alleviating the severity of other symptoms within that SC. This way, greater gains in a patients’ health-related quality of life (HRQoL) can be achieved and patient care can be simplified. The aims of this study are to identify (1) SC and their change over time in lung cancer patients undergoing radiotherapy (RT), (2) SC with the greatest impact on HRQoL, and (3) demographical, clinical and/or treatment-related predictors of SC. Materials and Methods Data were used from the REQUITE study: an international prospective cohort study including lung cancer patients receiving RT from 26 different hospitals and 8 countries. SC were identified based on patient-reported outcomes collected before RT(T1), at month 3(T2), and month 6(T3) after RT with the EORTC QLQ-C3O and the lung symptom questionnaire. A combination of the following machine learning techniques were used to identify symptom clusters at different time-points, to investigate the impact of the SC on HRQoL and to predict the SC, respectively: hierarchical agglomerative clustering, linear regression and random forest regression. To guarantee external validity of the prediction model, a first part of the data set was used to develop the prediction model, and a second part to validate the prediction model for unseen data. Results Data from 418, 341, and 299 lung cancer patients were analysed at T1, T2, and T3, respectively. Three SC were identified and remained stable over time: cluster 1 (fatigue, dyspnoea, physical and role functioning), cluster 2 (coughing blood, swallowing problems, nausea and diarrhoea), and cluster 3 (social, emotional and cognitive functioning). On T1 and T2, a 4th cluster was identified (general pain, chest pain and coughing). Cluster 1 was most common across all time points, followed by clusters 3, 4 and 2. At T1, cluster 3 had the greatest impact on overall HRQoL (34% explained variance) while cluster 1 had the greatest impact at T2 (39%) and T3 (50%). Two symptoms within cluster 1 (dyspnoea and physical functioning) could be moderately predicted at T2 with age and RT parameters (i.e. planned target volume, max. dose oesophagus and dose per fraction) being the greatest predictors. Conclusion Supportive care interventions for lung cancer patients undergoing RT must tackle 1 or 2 symptoms of the ‘fatigue, dyspnoea, physical and role functioning’ cluster because this SC is most common across time-points and has the greatest impact on the patients’ HRQoL. Furthermore, age and RT parameters should be taken into account to further tailor future interventions in lung cancer patients.