Metastatic breast cancer (MBC) is an incurable disease, wherein the available medications are primarily focused on deferring disease progression and symptom mitigation, thereby prolonging survival rate as well as preserving the quality of life (QoL) and psychological well-being [1,2]. The clinical advancements achieved in anticancer treatments have increased the survival rates of patients with MBC. The 5-year survival rate of patients with MBC is around 38% [2]. Notwithstanding, several studies [3-6] have shown that adherence to anticancer treatments is a critical issue in the disease trajectory of patients with breast cancer, especially regarding oral anticancer treatments (OATs), which are intensely demanding because patients are responsible for assuming medications according to the medical prescriptions, thereby increasing the risk of not appropriately taking the therapy [7].

From a theoretical perspective and according to the World Health Organization’s recommendations [8], medication adherence might be explained by a set of mutual and interconnected determinants incorporating (1) sociodemographic (eg, age, gender, socioeconomic status), psychocognitive, and social variables (eg, psychological well-being, social support); (2) disease and treatment-related characteristics (eg, cancer stage, prognosis, dosage, side effects); (3) attitudes, beliefs, and values; (4) health literacy and knowledge; and finally (5) health care system–related factors [7,9,10]. Accruing evidence have highlighted that patients with advanced cancer might report significant levels of nonadherence because in case of some cancers such as MBC, the patients have to change the type of treatments or dosage frequently and they experience a high level of fear of cancer spread [3]. For example, Yerrapragada and colleagues [6] reported nonadherence to tamoxifen in patients with MBC, ranging between 30% and 85% and further reducing over time. In addition, patients with MBC reported several barriers to the daily management of OAT, such as emotional (eg, worry, depression) and physical distress related to the side effects (eg, fatigue, weakness, sleep disturbance, emotional burden, pain) and lack of knowledge about their disease. Further, patients might experience a lack of control and a lack of perceived benefits during the disease pathway, and they may experience difficulty in managing therapy [3,6]. Other studies have shown that patients with MBC experience modest QoL due to the related treatment side effects, financial burden affecting therapy discontinuation, and a significant level of nonadherence [4,7]. Marshall and colleagues [4] observed that patients with MBC who frequently report treatment concerns rather than treatment benefits are less likely to be adherent to prescriptions, as well as patients who experience a more significant number and severity of side effects tend to have more medication worries that could negatively impact adherence.

Given the impact of nonadherence on clinical outcomes and the associated economic burden on the health care system [11,12], finding effective ways to increase treatment adherence is particularly relevant. Patients who adopt nonadherent behaviors need support in managing oral therapies and in overcoming individual and systemic barriers and roadblocks [5]. Notwithstanding, the dynamics influencing adherence to OATs are understudied in the cancer field [13], and a comprehensive model of the risk determinants of nonadherence is not currently available, particularly for targeted therapies and novel generations of hormonal therapy [14]. Besides, a shared definition of medication adherence and satisfactory assessment tools is not acknowledged, which vary in terms of accuracy and reliability [14-17], even in the explicit case of the equivalent treatment protocol and diagnosis [13,18]. Because of the direct costs of nonadherence (eg, survival rate, health-related QoL for patients with MBC) and the indirect costs for the health care system (eg, economic burden), it is mandatory to identify potential risk factors of nonadherence to OATs and to define personalized interventions supporting adherence during the clinical pathway in patients with MBC.

Consistently, defining, measuring, and developing a comprehensive model of medication adherence based on real-world data predictive models is a crucial clinical, psychological, social, and economic challenge. Machine learning has become an integral part of the health care industry—from biomedical research to the delivery of health care services. Compared to traditional statistical methods, machine learning provides numerous advantages such as increased flexibility, prediction accuracy, possibility of automation, and processing of big data. Prediction models for adherence have already been developed and tested in various scenarios [19-22]. If adequately reported, these models can help guide treatment decision-making, improve patient outcomes, and streamline perioperative health care management. Considering the complexity of medication nonadherence in patients with MBC, it is critical to identify patients at risk of nonadherence and carry out timely, precise, and tailored interventions to improve their adherence. Through machine learning models, it is possible to provide personalized prediction on medication adherence for a given patient, supporting adherence and performing a specific intervention [22,23].

A growing body of studies have underlined that eHealth technologies (eg, mobile apps, web-based solutions, wearables) might be helpful tools to foster patient management and engagement in clinical decisions during the cancer pathway [24]. Different web-based solutions based on educational and behavioral interventions have been developed for patients with breast cancer to foster medication adherence [25,26]. For example, the Multinational Association for Supportive Care in Cancer has developed an educational and teaching tool for patients with cancer receiving OATs that is composed of different educational sections aimed at assessing general patient knowledge about their treatment protocol and drug information (eg, side effects), skills in the management of therapy, possible strategies to manage nonadherence occurrences, and a specific questionnaire section to evaluate patient comprehension [26]. Moreover, Omaki and colleagues [27] have developed a patient decision aid named “My Healthy Choices” to foster adherence to pain treatments, assessing the environmental and personal risks and setting patient treatment priorities.

Nevertheless, no study has been conducted on patients with MBC to foster medication adherence to OATs through the clinical care pathway based on designing and developing a decision support system (DSS), integrating risk predictive models and educational and training tools. The information embedded in a DSS solution designed and developed according to the needs of patients with MBC might enable users to be better informed, develop more accurate expectations of the benefits and harms, and increase participation in the decision-making processes and medication adherence [28]. Evidence shows that, when implemented on web or mobile apps, DSS may support patients and physicians by improving adherence to medical treatments [27,29,30].

Drawing from the theoretical framework described above, we present and explain the study protocol of a prospective, randomized controlled study that is nested in a large-scale international project named “Enhancing Therapy Adherence Among Metastatic Breast Cancer Patients” (Pfizer 65080791) aimed to develop a predictive model of nonadherence and an associated DSS and guidelines to foster patients’ engagement and therapy adherence among patients with MBC concerning oral chemotherapy, endocrine therapy, supportive care, and the treatment of comorbidities. Consistently, the Pfizer Project (65080791) has been organized into 2 different studies to achieve the key goal. A retrospective study has been designed and a model has been developed to predict adherence to OATs, which use existing physiological, clinical, and QoL data available in the European Institute of Oncology (IEO; Milan, Italy). More in-detail, multimodal retrospective data have been retrieved from patient electronic health records by using natural language processing in a sample of 2750 patients with MBC (from 2010 to 2020). Data included in the analysis have been sociodemographic variables, diagnosis, biochemical and medical tests, procedures and medical history, treatment programs, treatment side effects, comorbidities, and familiarities. Concerning adherence to the treatment protocols, data included the following dimensions: initiation of the treatment, interruption of treatment, and skipped treatment doses. Furthermore, a prospective study is designed to assess the effectiveness of a DSS web-based solution and to enrich the predictive power of the machine learning model to forecast adherence behavior in patients with MBC. The tuning of the model permits adding additional predictors (personality traits, self-efficacy for coping with cancer, sense of coherence, pain, anxiety, depression, risk perception, and QoL) known to influence medication adherence behavior and that are not available retrospectively [10]. These data are used to improve the predictive power of the machine learning model and its capacity to profile patients’ adherence behaviors and to provide an individual risk value of nonadherence.

The primary end point was to assess the effectiveness of the DSS web-based solution and the machine learning web application in promoting adherence to OATs in a sample of 100 patients with MBC at 3 months. The adherence is evaluated using the number of pills taken divided by the number of pills prescribed. Further, the adherence is assessed using behavioral measures: the 8-item Morisky Medication Adherence Scale (MMAS-8) [31] and Adherence Attitude Inventory (AAI) [32].

The secondary end points were to identify clinical (comorbidities, presence of pain, tumor type, type of treatment), psychological (personality traits, anxiety, depression, self-efficacy for coping with cancer, sense of coherence, and risk perception), and QoL variables predicting patients’ adherence to OATs. These variables are used as predictors for evaluating nonadherence to OATs among patients with MBC and for enriching the preliminary version of a machine learning model developed in the retrospective study. Our initial machine learning models that are evaluated as an intervention in this study were built based on variables extracted from clinical notes by using natural language processing. Once the data collection for this study is complete, we will also use the data collected to improve our initial machine learning models. Data for the secondary end points are collected using the European Organization for Research and Treatment of Cancer QoL Questionnaire Core 30 (EORTC-QLQ-C30) [33,34], European Organization for Research and Treatment of Cancer QoL 23-item Breast Cancer-specific Questionnaire (EORTC-QLQ-BR23) [35], and the Brief Pain Inventory (BPI) [36]. To evaluate psychological variables, we used the State-Trait Anxiety Inventory (STAI-Y) [37,38], Beck Depression Inventory-II (BDI-II) [39,40], Big Five Inventory (BFI) [41], Brief Italian version of Cancer Behavior Inventory (CBI-B/I) [42,43], 13-item Sense of Coherence scale (SOC-13) [44], and risk perception (using 2 visual analogue scales) [45,46].

There are 3 assessment time points. At the baseline (T0), all patients fill a set of validated measures (MMAS-8, AAI, STAI-Y forms I and II, BDI-II, CBI-B/I, SOC-13, BPI, EORTC-QLQ-C30, EORTC-QLQ-BR23, BFI, and visual analogue scale). The expected time to complete all the given questionnaires at baseline is approximately 40 minutes. Access to DSS is given to the experimental group for 3 months. At T1 and T2, the following questionnaires are filled: MMAS-8, AAI, STAI-Y form I, BDI-II, CBI-B/I, EORTC-QLQ-C30, and EORTC-QLQ-BR23. Further, all patients have to fill a weekly adherence medication diary for 3 months. Variables that are not sensitive to change, for example, personality (BFI) and anxiety traits (STAI-Y form II) are collected only at T0. Each month, all participants receive a brief telephone interview in which they are monitored for adherence to the research protocol. Two psychologists perform the monthly telephone interview to monitor the filling out of the questionnaires and the medication adherence diary supporting patients with MBC in this task, barriers, and concerns related to the study participation. At T3 (3 months), all behavioral and psychological measures are filled, and an interview (online or vis-à-vis) is performed.

The sample size calculation was based on estimates for the effectiveness of TREAT as the primary outcome, assuming that the final analysis would be performed with a 2-sample 2-sided t test, a power of 96.4%, an α of .05, a standard deviation of 1.2, and a minimal difference in outcomes of 0.51 (effect size of 0.42) based on the patients’ self-reported satisfaction with the treatment decision [53]. Under these considerations, the final sample size was calculated to be 100 patients with MBC, with 50 individuals per group (50 patients with MBC exposed to DSS and 50 patients with MBC not exposed to the intervention). This sample size is expanded to 120 when a 20% attrition rate is used.

This study is compliant with the recommendations outlined in the Helsinki Declaration [57] and the Council for International Organizations of Medical Sciences guidelines [58], as well as with the principles of biomedical ethics reported in the Belmont report [59]. This study presents a fair balance between risks and benefits for study participants and future patients affected by the same condition. No physical risks directly related to participation in the research are expected. Although psychological risks are not expected, in case these raised from participation, psychologists responsible for the study will promptly intervene and take care of the patient. Regarding the benefits, given the impact of nonadherence on clinical outcomes, finding effective ways to increase treatment adherence is particularly relevant for the community of patients with MBC, thus playing a role in the improvement of the patient’s general well-being. The principle of self-determination is also respected. A devoted informed consent form is signed by participants before participation. Signing the informed consent form is preceded by a dialogic consent process necessary to ensure informed, voluntary, and awareness of participation in research. Regarding the respect for privacy, this study is performed according to the General Data Protection Regulation (Regulation EU, 2016/679) [60]. All data are collected in a pseudoanonymized form. Data are treated confidentially and used only by the collaborators in this study for scientific purposes related to what is stated in the research protocol. The retrospective study was approved by the IEO ethics committee (R1595/21-IEO 1704). The prospective study was approved by the IEO ethics committee (R1786/22-IEO 1907).

Adherence to anticancer treatment is fundamental in the clinical management of patients with MBC, as adherence can lead to a series of crucial clinical outcomes (eg, prolonged survival time, better monitoring and management of side effects, improvement of QoL) [1-6]. Studies on implementing machine learning to foster medication adherence as well as the development of shared DSSs are quite recent [23].

We believe that our study might achieve 3 milestones in the clinical care of patients with MBC with regard to adherence to OATs. Concerning the primary end point, implementing a shared and integrated DSS web-based solution might be an essential strategy for patients with MBC to enhance their health knowledge and understanding of medication nonadherence issues and to learn behavioral strategies to overcome individual, disease-related, and organizational barriers impacting adherence behaviors. This expected result is coherent with the results of other studies on patient decision tools, highlighting that such engines might improve patients’ participation and health knowledge in the care pathway [61]. Regarding the secondary end point, this prospective study will contribute to refining the initial risk-predictive model of adherence attained from the retrospective study, informing about specific (inner and external) predictors of adherence behavior during the MBC care pathway that were not considered in the preliminary version of the model. This should provide a more comprehensive and systematic definition of medication adherence for OATs in patients with MBC, defining an interrelated and evidence-based list of predictors. Data will define a final risk-predictive model of adherence and enable the identification of specific patient populations at risk of nonadherence. The earlier identification of patients with poor adherence using our machine learning web application will permit the development of tailored psychological and behavioral interventions to foster adherence by dealing with individual barriers and needs according to the risk profile. Finally, the information retrieved by the risk-predictive models might support oncologists’ treatment decisions, allowing them to better understand the hindered reasons behind a complex adherence trajectory.

Despite the key contributions of this study in the field of the medication adherence to OATs among patients with MBC, some limitations have to be acknowledged. First, the questionnaires used to refine our risk-predictive model might cause cognitive burden and general fatigue among participants. However, a monthly telephone interview has been introduced as a mitigation risk strategy. The interview is aimed to support participants in filling out the questionnaires and medication adherence diary and to manage barriers and concerns related to this study. Related to this aspect, the second limitation concerns the recall bias that might be generated using self-report measures. Further, self-report measures might overvalue adherence compared to the other prospective methods such as the Medication Event Monitoring System. However, the weekly medication diary method, in which patients have to report a detailed and systematic description of pill counts, side effects, and barriers related to the management of the therapy and missed doses, can aid in providing a more comprehensive picture of medication adherence. Moreover, as highlighted by Shah and colleagues [15], the diary has a higher accuracy and minor recall bias and permits the evaluation of subjective information (eg, emotions, thoughts, worries, behaviors, expectations) related to the medication coherently with the processual model of the adherence used in this research protocol [8]. Third, a monthly follow-up and a 3-month end point have been established, considering the type of diagnosis, high variation in the prognosis, and shifting in the lines of treatment due to cancer progression or declining physical functioning. This short end point might affect a deep understanding of the efficacy of our DSS in the long term. However, previous studies have used correspondent follow-ups and end points [13]. Fourth, as observed by other studies on patients with MBC [14], an additional risk that should be considered is related to the overestimated adherence rate. Patients participating in clinical trials are highly motivated, possibly leading to greater adherence to the OATs prescribed. Fifth, the initial version of the machine learning web application was not able to directly predict patient adherence. Instead, it predicted adherence risk factors, including short-term and long-term side effects, as well as physical status and comorbid conditions. This is a limitation of the data we used to build this initial version of the model, as it was not possible to reliably identify adherence patterns. However, predicting adherence risk factors can still provide insights that can assist in improving OAT adherence during shared decision-making sessions between physicians and patients.

Considering the poor evidence on OATs and the need to develop validated instruments to evaluate medication adherence to improve patient clinical outcomes [14,16], the development of integrated and shared DSSs able to foster adherence behaviors and to profile patient adherence risks might have key impacts on clinical practice and patient health-related QoL, thereby enabling the early identification of high-risk populations and enriching knowledge about the implementation of machine learning models in clinical practice.