Predicting onset and continuity of patient-reported symptoms in cancer patients undergoing immune checkpoint inhibitor (ICI) therapies using machine learning

Abstract

Abstract Background In recent years, ICI therapies have become standard-of-care treatments in several malignancies. ICIs introduced novel toxicities, which can arise from various organ systems or at any time point even after treatment discontinuation. The toxicities can be life threatening, but most are reversible if detected and treated early and therefore, early detection could result in improved safety profile of treatment and better Health related Quality-of-Life. Electronic collection of patient reported symptoms (ePRO) could be used in development of machine learning (ML) based prediction models to enable earlier detection of toxicities compared to traditional follow-up. Methods The training dataset consists of ePRO data of ICI-treated patients collected using Kaiku Health platform, including 21 744 reported symptoms from 72 ICI patients. The ML models were built with an established classification algorithm extreme gradient boosting and trained for 14 symptoms related to ICI toxicities. The validation dataset consisted of 16 884 reported symptoms from 67 cancer patients receiving ICIs and followed with the ePRO tool. Results ML models were able to predict the onset and continuation of 14 symptoms that may indicate the development of ICI toxicities. The model performance assessed using area under curve (AUC), a common performance metric for ML models, is best for dyspnea (0.96) and lowest for headache (0.77). Generally, the models are performing well with an average AUC of 0.86, as shown in the table. Table . 50P AUC values for symptoms related to ICI toxicities from the validation dataset Symptoms AUC Dizziness 0,88 Itching 0,84 Fever 0,90 Diarrhoea 0,78 Stomach pain 0,86 Nausea 0,82 Fatigue 0,90 Rash 0,84 Decreased appetite 0,84 Cough 0,90 Dyspnea 0,96 Joint pain 0,92 Headache 0,77 Chest pain 0,88 Conclusion ML based modeling of ePRO data on ICI treated cancer patients is feasible in predicting onset and continuation of symptoms related to ICI toxicities. The study suggest that ML based approaches could be used in early detection of toxicities. The results should be validated with a dataset collected in a prospective clinical trial. Legal entity responsible for the study Kaiku Health Oy. Funding Oulu University and Finnish Cancer Society. Disclosure J. Ekstrom: Full / Part-time employment: Kaiku Health Oy. H. Virtanen: Full / Part-time employment: Kaiku Health Oy; Shareholder / Stockholder / Stock options: Kaiku Health Oy. J.P. Koivunen: Advisory / Consultancy: Kaiku Health Oy. All other authors have declared no conflicts of interest.