Andreas Trojan ^1* Emanuele Laurenzi ^2* Sven Roth ^3* Michael Kiessling ^3* Ziad Atassi ^4* Yannick Kadvany ^4* Meinrad Mannhart ^4* Hans F. Witschel ² Stephan Jüngling ^2* Gerd A. Kullak-Ublick ^3,5 Christian Jackisch ^6*

• ¹ Institute of Pharmacology and Toxicology, Faculty of Medicine, University of Zurich, Zurich, Zürich, Switzerland
• ² University of Applied Sciences and Arts Northwestern Switzerland, Windisch, Switzerland
• ³ Pharmakologie und Toxikologie, University of Zurich, Zürich, Switzerland
• ⁴ Breast-Center Zurich, Zurich, Switzerland
• ⁵ University Hospital Zürich, Zurich, Zürich, Switzerland
• ⁶ Sana Klinikum Offenbach, Offenbach, Germany

Background: Capturing and analyzing patient-reported outcomes (ePRO) can promote the early detection of potential adverse events and may be supported by machine learning (ML) for reduction of adverse events and unplanned admissions. Objective: We aimed to create an Early Warning System (EWS) to predict situations where supportive interventions become necessary to prevent unplanned visits. For this, dynamically collected standardized ePRO data were analyzed. Information on well-being, vital parameters and medication were also considered for establishing a hybrid ML model. Given the limitations of highly imbalanced datasets (where only very few adverse events are present) and the limitations of humans in overseeing all possible causes of such events, we hypothesize that it should be possible to combine both to partially overcome these limitations. The prediction of unplanned visits was achieved by employing a white-box ML algorithm (i.e., rule learner), which learned human-interpretable and -modifiable rules from ePRO data. Rules were evaluated and ranked based on cost that preferred sensitivity over precision. The best rules were reviewed by two oncological experts for plausibility and extended with additional conditions. Results: From a cohort of 214 patients and more than 16'000 data entries, the machine-learned rule set achieved a recall of 19% on the entire dataset and a precision of 5%. We compared this performance to a set of conditions that a human expert had defined. This "human baseline" did not discover any of the adverse events recorded in our dataset (precision and recall 0%). Despite the poor initial ML results, the involved medical experts understood the rules and felt capable of suggesting modifications to increase their precision. Modifications of rules included adding or tightening conditions to make them less sensitive or changing the rule consequences. We can conclude that it is possible to apply ML to inspire human experts. While humans seem to lack the ability to define such rules without support, they are capable of modifying the rules to increase their precision and generalizability. Conclusions: Learning rules from dynamic ePRO datasets may be used to assist human experts in establishing an early warning system for cancer patients in outpatient settings.