BACKGROUND
Recently the use of smartphone apps has been implemented in clinical trials with cancer patients undergoing systemic treatment. This approach has demonstrated the feasibility of those apps in the early detection of symptoms and therapy side effects. Early detection has the positive effects of (a) enabling the timely adaption of current therapies and (b) reducing the number of acute admissions.
OBJECTIVE
This study aimed to create an Early Warning System (EWS) for the improvement of digital monitoring of cancer patients. The proposed EWS relies on “electronically captured patient reported outcomes” (ePROs) to predict situations where supportive interventions are necessary to prevent unplanned visits and acute admissions. The utilized ePROs contain information on well-being, symptoms, vital parameters, and medication, which is considered reliable if reported in a standardized and structured manner.
METHODS
The prediction of unplanned visits was achieved by employing a white-box Machine Learning (ML) algorithm (i.e., rule learner), which learned rules from patient data (i.e., ePROs) that had been captured by means of a smartphone app. The learned rules indicated situations where patients had attended unplanned visits and hence were captured as alert triggers in the EWS. Each rule was evaluated based on a cost matrix, where false negatives (FNs) were assigned higher cost than false positives (FPs; i.e., false alarms). False alarms are not considered as bad as false negatives as it is less harming to have an EWS raising a warning when it is not necessary rather than not raising a warning when it is necessary. Next, the rules were ranked according to the cost, and priority was given to the least expensive ones. Finally, the rules with higher priority were reviewed by oncological experts to check their plausibility and extend them with additional conditions.
RESULTS
From a cohort of 214 patients and a total of more than 16,000 data entries overall, the machine-learned rule set achieved a recall of 28% on the entire dataset and a precision of 15%. While modification by medical experts did not improve these values on the given data set, they expressed confidence that they had understood and were able to make sense of the rules and that the modified rules could be expected to generalize better to unseen data. Furthermore, the involvement of experts allowed to define rule consequences, i.e., actions to be recommended to patients or caregivers that are not necessarily visits to physicians or hospitals. When we compared the effectiveness of the learned rules to rules that our medical experts had.
CONCLUSIONS
The proposed approach showed the feasibility of learning rules for the purpose of an early warning system from ePRO datasets. New rules can be learned as the ePRO dataset increases. Our approach sets the basis for the creation of an Early Warning System for the homecare treatment of cancer patients. Next, we will make the ML component more sophisticated by considering more data and additional data types from IoT devices (Bluetooth® wearables). Based on these rules, we will investigate the types of recommendations that we can provide to patients and to oncologists. The recommendations will then be associated with the warnings, thus supplementing the Early Warning System.