1839P - Dynamic reporting of treatment related symptoms via ePROs can reversely identify the type of underlying cancer

Background

Digital symptom reporting through cancer patients (ePRO) undergoing systemic treatment has demonstrated early detection of symptoms, equivalent side effects regarding similar drugs, reduction of unplanned admissions, and machine learning (ML) supported prediction when patients will require uplanned admissions. We examined if dynamic reporting of treatment related symptoms via ePROs can reversely identify the type of underlying cancer.

Methods

226 patients on treatment had self-reported on presence and severity (according to CTCAE) of more than 90 available symptoms via the medidux app (formerly consilium care). For a balanced analysis data from 25 patients treated for breast cancer, 19 lung cancer, 16 colon cancer, 12 lymphoma and 7 prostate cancer, respectively, were used. Patients` symptoms over the entire study period were aggregated by counting the days on which a particular symptom was reported. Thus, each patient was represented by a vector of symptoms indicating how often the given symptom occurred. A human-interpretable ML logistic regression model was applied to predict the primary tumor of the patient from his/her respective symptom vector. All symptoms with positive coefficient above a certain threshold (0.1) were collected and then graphically displayed for association between symptoms and cancer type.

Results

The ML model was not able to recognize the prostate and blood-lymph patients in retrospect since their number was too small. Analysis for the three remaining cancer types revealed a mean area under the curve (AUC) score of 0.72 (breast cancer AUC 0.74, CI: 0.62–0.85; gut cancer AUC 0.78, CI: 0.66–0.89; lung cancer AUC 0.63, CI: 0.50–0.77). Results indicate that ML performs “fair” and significantly better than random guessing (which would result in AUC = 0.5) for the reverse identification of the underlying cancer upon ePRO reporting from patients.

Conclusions

Cloud aggregation of ePROs and ML harbor the cost-effective potential in identifying specific side effects and the underlying cancer for which patients receive systemic treatment. Whether associations can be made for dynamic changes of reported symptoms and adherence to oral medication shall be explored in prospective decentralized clinical trials.

Clinical trial identification

NCT03578731.

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

A. Trojan: Financial Interests, Personal, Stocks or ownership, Chief Medical Officer: mobile Health AG. All other authors have declared no conflicts of interest.