Supervised machine learning applied in nursing notes for identifying childhood cancer patients' need for psychosocial support

Akseli Reunamo^1*Hans Moen²Sanna Salanterä^3,4Päivi M Lähteenmäki⁵

• ¹Department of Computing, Faculty of Technology, University of Turku, Turku, Finland
• ²Department of Computer Science, School of Science, Aalto University, Otakaari, Ostrobothnia, Finland
• ³Department of Nursing Science, Faculty of Medicine, University of Turku, Turku, Finland
• ⁴Nursing Administration, Turku University Hospital, Turku, Finland
• ⁵Department of Paediatrics and Adolescent Medicine, Turku University Hospital, Turku, Finland

IntroductionChildhood cancer survivors have a higher risk for mental health and adaptive problems compared, e.g., to their siblings. Assessing the need for psychosocial support is essential for prevention. This project aimed to investigate the use of supervised machine learning in the form of text classification in identifying childhood cancer patients needing psychosocial support from nursing notes when at least one year had passed from their cancer diagnosis.MethodsWe evaluated three well-known machine learning-based models to recognize patients who had outpatient clinic reservations in the mental health-related care units from free-text nursing notes of 1672 patients. For model training, the patients were children diagnosed with diabetes mellitus or cancer, while evaluation of the model was done on the patients diagnosed with cancer. A stratified five-fold nested cross-validation was used. We designed this as a binary classification task, with the labels: no support (0) or psychosocial support (1) was needed. Patients with the latter were identified by having an outpatient appointment reservation in a mental health-related care unit at least one year after their primary diagnosis.ResultsThe Random Forest classification model trained on both cancer and diabetes patients performed best for the cancer patient population in three times repeated nested-cross validation with 0.798 mean area under the receiver operating characteristics curve and was better with 99% probability (credibility interval - 0.2840, - 0.0422) than the neural network-based model using only cancer patients in training when comparing all classifiers pairwise by Bayesian Correlated t-test. ConclusionsUsing machine learning to predict childhood cancer patients needing psychosocial support was possible using nursing notes with a good area under the receiver operating characteristics curve. The reported experiment indicates that machine learning may assist in identifying patients likely to need mental health-related support later in life.