A 1-arm, multicenter, prospective clinical trial (“Follow-up of Cancer Patients Receiving Chemotherapy or Targeted Therapy by Electronic Patient Reported Outcomes-tool” [ECHO] 7/2019-1/2021; NCT04081558) investigated the use of a novel ePRO tool, Kaiku Health, in cancer care. The most important patient inclusion criteria were colorectal cancer (CRC), a plan to receive oxaliplatin-based chemotherapy as an adjuvant therapy or in the first- or second-line setting of advanced disease, age ≥18 years, an Eastern Cooperative Oncology Group [ECOG] performance score of 0 to 2, and internet access. A decision support tool consisting of a digital symptom-monitoring tool with an urgency algorithm, laboratory values, and integration of a treatment schedule to automate chemotherapy-cycle prescribing was created for the trial.

The Kaiku Health ePRO follow-up module consists of a questionnaire that assess both the presence and severity of symptoms related to typical side effects of chemotherapy (blood in urine, dysuria, eye symptoms [decreased vision or other symptoms], peripheral sensory neuropathy, pain, constipation, cough, decreased appetite, diarrhea, fatigue, fever, mouth sores, nausea, rash or skin changes, shortness of breath, and vomiting). The tool also allows recording the presence of a symptom and a severity algorithm that grades the symptom according to National Cancer Institute Common Terminology Criteria for Adverse Events (NCI-CTCAE). The severity algorithm triggered an email alert to the study physician in the care unit within preset limits (the presence of a grade 3 or higher symptom or a rise in symptom severity from grade 0 to 2). The patients were informed that the ePRO follow-up was intended only for nonurgent communication, and, in critical matters, patients were advised to contact emergency care. In addition, patients received tailored, evidence-based, personalized self-care advice according to electronically reported symptoms and their grade. The tool also included a messaging option through which patients could communicate with their care team directly.

Prior to chemotherapy infusion, peripheral blood samples were taken to evaluate bone marrow, kidney, and liver function. The laboratory values include counts for red and white cells, thrombocytes, alanine and aspartate aminotransferases, bilirubin, and creatinine. Based on the summary of product characteristics of the chemotherapeutic agents used by the patient, prespecified laboratory values were set to assess the fitness of the patient for a new course of cancer medical therapy.

We developed and added a novel element to the Kaiku Health remote symptom-monitoring tool: in addition to collecting and grading ePROs, it automatically collected laboratory values via a novel interface (MyLab). The timing of the ePRO symptom questionnaires and laboratory value collection was automated according to chemotherapy cycles through another interface with a treatment-appointment reservation module (Oberon). The decision support tool analyzed the completed symptom questionnaires and preceding laboratory values and compared the results to a set of institutional reference values. The decision support tool provided a 3-tier recommendation to the care team on the initiation of the preplanned chemotherapy cycle. The 3-tier recommendations included green (meaning “go”), yellow (“evaluate and possibly postpone”), or red (“postpone”). A recommendation of go versus no-go and potential treatment visit details were sent to the patient through the Kaiku Health tool. With a no-go decision, an automated ePRO questionnaire and laboratory value collection for the rescheduled chemotherapy cycle were initiated based on the reservation module interface.

The study was approved by the ethics committee of the Northern Ostrobothnia Hospital District (273/2017) and registered in an international clinical study registry (ClinicalTrials.gov NCT04081558). The study was carried out in accordance with the principles of the Declaration of Helsinki and Good Clinical Practice. All patients provided written informed consent. Study participants were not compensated for their participation in the trial.

Before a new chemotherapy cycle, a patient’s laboratory values and symptoms were evaluated, for example, for indications of treatable chemotherapy side effects, excess chemotherapy toxicity, and cancer progression. To streamline this work process, increase the comprehensiveness and repeatability of the precycle evaluation, and automate duties, we developed a decision support tool. The designed tool had 6 different elements that aimed to increase the efficiency and quality of cancer care (Figure 1; Multimedia Appendix 1).

The first element was a scheduling module that synchronized the sending of symptom questionnaires to patients via Kaiku Health and laboratory value collection from hospital registries via an interface. The symptom module used ePRO-based (Kaiku Health) grading (NCI-CTCAE) for 15 core symptoms. The laboratory-value module collected specific central values via a software interface (Figure 2A). The evaluation module compared the symptoms and laboratory values to values in a set range and assigned a recommendation for the chemotherapy cycle in 1 of 3 tiers (Figure 2B). For the tier-1 (green/go) recommendation, laboratory values needed to be within a set range and all the symptoms needed to be at grade ≤1; thus, the nurse made the planned chemotherapy order to the hospital pharmacy, and an SMS text message with information on the timetable of drug infusion was sent through Kaiku Health. For the tier-2 (yellow/evaluate) recommendation, laboratory values had to be within a set range and ePROs had to include ≥1 symptom that was grade ≥2 (if they were not already at baseline); the nurse individually assessed the symptoms. If there was a need for cancer drug modification, typically dose reduction or supportive medications, the study physician was consulted as to whether the chemotherapy cycle could be initiated or modified, or if supportive medication should be initiated or modified. In the case of a “go” decision, confirmation of the treatment time was sent via Kaiku Health. In the tier-3 (red/hold) scenario, laboratory values were out of the set range, and the nurse postponed the chemotherapy cycle and sent the new infusion date through Kaiku Health. In the case of a postponed chemotherapy cycle, the scheduling module programmed a new collection of symptoms and laboratory values according to the new infusion date.

The care team could always override the recommendation of the decision support system. In the case of missing ePRO or laboratory data (or if the data were >3 days old) before the chemotherapy cycle, the decision support tool gave a tier-2 (yellow/evaluate) recommendation; thus, manual evaluation had to be done prior to prescribing the treatment. The SMS messaging module was used to inform the patient about the recommendation and potential treatment decision. Furthermore, the messaging module enabled secure messaging between the patient and care team. The symptom support module also provided the patient with personalized guidance for self-care based on the presence of specific symptoms and their grade (Figure 2C).

Next, we analyzed the performance of the decision support tool based on data collected in a prospective, multicenter, single-arm clinical trial (ECHO; NCT04081558). Patient recruitment for the prospective cohort took place from July 2019 to January 2021 at the Oulu University Hospital and Vaasa Central Hospital oncology units. The study dataset included patients with CRC (n=43) in the ECHO trial who gave informed consent and were treated with doublet or triplet chemotherapy in an adjuvant or metastatic setting. During the study, a total of 339 chemotherapy cycles were prescribed. Altogether, 843 ePRO symptom questionnaires were completed during the ECHO study. Of the recorded 15 symptoms, fatigue (n=446, 52.9%) and peripheral neuropathy (n=429, 50.9%) were reported most often, while 137 grade 3 or 4 symptoms were recorded, of which diarrhea (n=5, 4%) and peripheral neuropathy (n=4, 3%) were the most common. Since the functioning of the Kaiku Health symptom tool has been evaluated in multiple previous trials [16-18], we focused here on the performance of the scheduling and evaluation modules.

The function of the scheduling module was assessed by evaluating the percentage of chemotherapy cycles with sent and completed symptom questionnaires. There were no false-positive or false-negative decision support tool recommendations. For over two-thirds (n=262, 77%) of the cycles, ePRO questionnaire data were available. Preceding a new chemotherapy cycle, 65% (n=221) of cycles had symptom questionnaires graded at ≤1, while 67% (n=228) of the cycles had laboratory values in the preset range. Evaluation module recommendations were tier 1 (green/go) in 145 (42.8%) of the cycles, tier 2 (yellow/evaluate) in 83 (25%), and tier 3 (red/hold) in 111 (32.7%), while 61% of the patients (n=27) needed a phone call before any of the planned chemotherapy cycles (Table 1).

The experience of the health care professionals (HCPs) with the decision support system integration into the care pathway was evaluated via interviews (n=4) that explored 5 topics: improved workflow, improved care, use of resources, challenges, and points for improvement. The HCPs appreciated the improved workflow, faster patient evaluation, and direct messaging option. The laboratory module originally collected data only daily, which was considered too infrequent by the HCPs, and data collection was reprogrammed to occur multiple times a day, which improved the usability of the tool. Some inconveniences were experienced with the scheduling module for patients with Baxter pump–based treatments, for whom 2 subsequent treatment reservations were made at once (treatment and pump removal); this led to incorrect scheduling of the symptom questionnaires. The HCPs felt that the messaging option should include read notifications, which would enable verification that the patient was aware of the chemotherapy schedule.

We present the results of a feasibility study investigating an integrated approach with an automated decision support system for chemotherapy-cycle prescription. We show that the investigated multidimensional integrated system is feasible and can provide clinically meaningful decision support. The approach could decrease the HCP workload and increase the quality and safety of cancer care.

In recent years, the use of ePRO symptom data has gained a great deal of interest in the oncology field. Previous studies have shown that ePRO monitoring can reduce emergency room visits, improve QoL, and, more importantly, increase survival [4-6]. Integration of an ePRO system into electronic patient records is considered essential for large-scale adoption of the approach. Timely integration of data from multiple electronic health care systems into ePROs could further enhance the clinical value of the captured data and facilitate decision support.

AI-based approaches can be superior for automatic detection, classification, and interpretation using big data. However, AIs can support improper decisions if they are presented with data beyond their original training set data. One could speculate that AI-based decision support systems could surpass our approach with a simple comparison to reference values. However, chemotherapy has a narrow therapeutic window and relying on prescription decisions made by AI alone may not satisfy ethical requirements without wide-scale training and validation [19-21].

A fast track for AI integration in clinical practice could lie in the use of large language models (LLMs). Methodological developments in LLMs such as in-context learning and retrieval-augmented generation could improve LLM performance and reduce hallucinations, consequently making the use of LLMs possible in clinical practice. It has been suggested that the recent development of advanced LLMs could improve patient care across several areas, such as clinical decision support or helping to answer patients’ questions, and could even improve patient outcomes by identifying social determinants of health and reducing health disparities [22-24].

In this study, we investigated a next-generation approach aiming to streamline and automate the prescription of chemotherapy cycles. We identified symptom profiles and laboratory values as central decision mediators for safe chemotherapy prescription. We hypothesized that symptom profiles could be collected via ePROs in a more comprehensive manner and used as input for a decision support tool that would also collect laboratory data and schedule collections based on the expected cycle date in an automated fashion. Furthermore, our approach included an expanded ePRO module with an integrated urgency algorithm, symptom-based personalized feedback and self-care advice, and a safe messaging platform.

We were able to show that the investigated platform is technically feasible, with good user experiences among patients and HCPs. More importantly, we found that automated “go” decisions for new chemotherapy cycles were made in almost half of the chemotherapy cycles. It is likely that this proportion could be increased by focusing only on core symptoms, instead of the total of 17 symptoms that were recorded in this study. We speculate that the investigated tool would decrease the workload for HCPs and enable allocation of human resources to more value-creating duties. We strongly believe that the computerized comparison of data to a set reference is more precise and error-free than human assessment, even though our study did not provide any metrics on the specific matter.

Nevertheless, a few modifications to the platform were suggested by end users. First, in the testing phase, the frequency of laboratory data collection was already considered too low; thus, the laboratory module was reprogrammed to search for data multiple times a day, which greatly improved the usability of the subtool. Second, the HCPs felt that the messaging option regarding the schedule of the planned chemotherapy cycle should include read notifications, which would enable verification that the patient was aware of the coming cycle. Furthermore, due to the low number of study participants and the observational nature of the study, the results are to be interpreted with caution, and the generalizability of the feasibility of the studied platform is unclear.

Due to the feasibility design of the study, the results are generally hypothesis generating. Since the study investigated a complex technical tool, initial testing with a feasibility approach is therefore warranted before progressing to a larger clinical study. The integration of an LLM-based module for interactivity, such as a chat box, could enhance the supportive role of the developed tool by providing educational elements to empower patients as well as give guidance on milder-grade chemotherapy toxicity management on a more individual level. Furthermore, the rather large proportion of symptom questionnaires with grade 2 symptoms might be decreased by fine-tuning the symptom evaluation, possibly by using AI-based models to mimic diagnostic pattern recognition to identify red-flag symptoms in parallel with laboratory value assessment. To conclude, the results of our trial are promising and lay the groundwork for further studies as well as development of the platform. A randomized study design would be required to fully validate the clinical utility and cost-effectiveness of the system.