Innovations in Research and Clinical Care Using Patient-Generated Health Data

Abstract

Patient-generated health data (PGHD), or health-related data gathered from patients to help address a health concern, is increasingly used to make regulatory decisions and evaluate quality of care. PGHD includes self-reported health and treatment histories, patient-reported outcomes (PROs), and biometric sensor data. Advances in wireless technology, smartphones, and the Internet of Things (IoT) have facilitated new ways to collect PGHD during clinic visits and in daily life. The goal of the current review is to provide an overview of the current clinical, regulatory, technological, and analytic landscape as it relates to PGHD in oncology research and care. The review begins with a rationale for PGHD as described by the U.S. Food and Drug Administration, Institute of Medicine, and other regulatory and scientific organizations. The evidence base for clinic-based and remote symptom monitoring using PGHD is described, with an emphasis on PROs. An overview is presented of current approaches to digital phenotyping, or device-based, real-time assessment of biometric, behavioral, self-report, and performance data. Analytic opportunities regarding PGHD are envisioned in the context of big data and artificial intelligence in medicine. Finally, challenges and solutions for integration of PGHD into clinical care are presented. Challenges include EMR integration of PROs and biometric data, analysis of large and complex biometric datasets, and potential clinic workflow redesign. In addition, there is currently more limited evidence for use of biometric data relative to PROs. Despite these challenges, the potential benefits of PGHD make it increasingly likely to be integrated into oncology research and clinical care.

The past decade has seen remarkable progress in the translation of biological discoveries into new cancer treatments, such as targeted therapies, immune checkpoint inhibitors, and adaptive cellular therapies. The effect of each new treatment is incremental but cumulatively their impact on survival rates, together with previous advances in early detection, has been unprecedented. Although diagnosis and treatment can still be a lonely and fearful experience for patients, in many cases cancer is becoming a chronic condition rather than a fatal disease. The result is that maintenance of quality of life has become an increasingly important clinical goal.

Quality of life is one aspect of patient-generated health data (PGHD), or health-related data gathered from patients to help address a health concern.^1,2 PGHD includes self-reported health and treatment histories, patient-reported outcomes (PROs), and biometric data (see Figure 1). PROs are defined by the U.S. Food and Drug Administration (FDA) as “reports of the status of a patient’s health condition that come directly from the patient, without interpretation of the patient’s response by a clinician or anyone else.”³ PROs can be categorized as disease-related symptoms, side effects of treatment, and quality of life (i.e., how symptoms and side effects impact daily functioning).⁴ Whereas self-reported health and treatment histories are well-established in the clinical setting and PROs in the research setting, the use of biometric PGHD data is still in early development. Biometric data can include passively-collected data from wearable sensors (e.g., a physical activity tracker) as well as data actively collected by patients through other instruments (e.g., a wireless blood pressure cuff). Advances in wireless technology, smartphones, and the Internet of Things (IoT) have facilitated new ways to collect PGHD during point-of-care clinic visits and in routine daily life. The current article will review opportunities and challenges of PGHD to inform regulatory decisions and cancer care delivery.

Figure 1.

Examples of patient-generated health data as defined by the U.S. Department of Health and Human Services.¹

The Need for Patient-Generated Health Data in Cancer Care

Growing interest in PGHD in oncology both reflects and reinforces an increasing role for patient advocates in directing regulatory priorities.⁵ In 2009 the U.S. FDA released a draft guidance document encouraging patient-focused drug development to ensure that the patient experience is sufficiently represented in benefit-risk assessment.³ Patient-focused drug development has primarily taken the form of collection of PROs as secondary outcomes in Phase III clinical trials,³ although adherence to guidelines for implementation has been sub-optimal.⁶ With the exception of one trial that collected home-based blood pressure readings,⁷ remote collection of biometric data has been limited. Nevertheless, collection of PROs is important because there are extensive data that suggest PROs provide information that is complementary to but different than clinician-rated adverse events. While clinicians’ ratings of adverse events are informed by medical knowledge, they tend to significantly underestimate patients’ reports of symptomatology. Severe symptomatology has been shown to be under-reported on clinical trials by up to 76%.^8,9 In contrast, data suggest that PROs are more sensitive to treatment-related differences in toxicity than clinician-rated adverse events.¹⁰ Notably, patients treated on clinical trials tend to be younger, healthier, and have higher socioeconomic status than patients treated outside of clinical trials in the community setting.^11,12 Thus, adverse events reported on clinical trials may not be generalizable to patients receiving the same treatment as standard of care. In contrast, PROs give a voice to patients.^13,14 Without PRO data from high-quality studies,¹⁵ patients may instead rely on anecdotal information from the internet about what to expect for a given disease and treatment.¹⁶ This anecdotal information may be inaccurate or ungeneralizable.¹⁶ PRO data are particularly important when patients must decide between two or more treatments that demonstrate similar or modest benefits regarding survival. PRO data can inform decision making in this situation by providing insight into quality of life, such as the ability to maintain one’s roles and responsibilities during treatment (e.g., continuing to work), which can also have significant emotional and financial benefits.¹⁶

PGHD also reflects increased awareness of the importance of proactive symptom management in high-quality cancer care.^13,17–19 Proactive symptom management is part of a larger trend recommended by the Institute of Medicine (IOM) to engage patients to improve quality of care. As recently as 1999, PROs were framed in terms of understanding tradeoffs between quantity and quality of life.²⁰ However, a series of studies published starting in 2010 demonstrated that early palliative care improved survival by an average of 4.6 months in patients with advanced cancer.^21–23 These and other studies showed that early palliative care also improved quality of life and reduced distress in both patients and caregivers.^21,23–25 Findings were extended in a recent high-visibility study demonstrating that clinic-based symptom monitoring and management improved quality of life and extended survival by 5 months in cancer patients treated with chemotherapy, perhaps because patients received chemotherapy longer.^26,27 Notably, the survival benefits of symptom management compare favorably to U.S. FDA approved anti-cancer agents approved between 2009 and 2013, which demonstrated a median survival benefit of 2.7 months.²⁸ Thus, it has become evident that improving quality of life can also lengthen quantity of life.

Opportunities abound to improve symptom management. Symptomatology has been chronicled most comprehensively by Cancer Care Ontario through systematic administration of the Edmonton Symptom Assessment Scale (ESAS) to oncology patients since 2007.²⁹ In a study of 120,745 patients within 12 months of diagnosis, the most commonly-endorsed moderate to severe symptoms were tiredness (59%), low overall well-being (55%), anxiety (44%), lack of appetite (43%), and pain (37%).²⁹ A diagnosis of respiratory or oropharyngeal cancer, younger age, female sex, lower income, greater comorbidities, and urban residence were associated with significantly higher odds of elevated symptom burden.²⁹ These findings are consistent with meta-analyses estimating the prevalence of common patient-reported symptoms. For example, the prevalence of pain is estimated to be 55% during anticancer treatment; 39% after curative treatment; 66% in advanced, metastatic, or terminal disease; and 50.7% in all cancer stages.³⁰ The prevalence of fatigue is estimated to be 14–27% among breast cancer survivors, 78% among older patients receiving palliative care, and 7% in the general population.^31–33 The prevalence of depression is estimated to be 27% during treatment, 21% during the year after diagnosis, 15% after one year following diagnosis, and 12% after two or more years following diagnosis.^34,35 In contrast, depression in non-cancer controls assessed similarly is estimated to be 10%.³⁵ The prevalence of anxiety is estimated to be 18% in long-term cancer survivors versus 13% in non-cancer controls assessed similarly.³⁵ A large study found the prevalence of distress to be 46% across the cancer continuum in 55 cancer centers in North America,³⁶ similar to rates reported in Europe.³⁷ Thus, while symptoms of depression and anxiety returned to normative levels in long-term survivorship, the prevalence of pain and fatigue remain high across the survivorship continuum. Rates of these symptoms demonstrate a significant unmet need for symptom monitoring and management, particularly during active treatment and in patients with advanced cancer.

Symptomatology tends to be under-recognized in clinical care, which may contribute to its high prevalence. Outside of the context of oncology, it has been noted that 31% of patients reporting chest pain, 38% reporting dyspnea, and 45% reporting cough on a clinical visit information form did not have documentation of the symptom in electronic medical record (EMR).³⁸ A large multicenter study found that oncologists underestimated rates of patients’ symptoms, particularly among patients with a poor Karnofsky Performance Status or poor Mini Mental State score, or those who were hospitalized, recently diagnosed, or undergoing opioid titration.³⁹ Even on a palliative care inpatient unit, nurses’ ratings of patients’ symptoms were not significantly correlated with patients’ own ratings of their symptoms.⁴⁰ Providers’ perceptions of symptoms differ not only from patients but also from other providers. In a study of symptoms assessed by pairs of providers approximately one hour apart, intraclass correlations were fair to poor across a range of symptoms including neuropathy, dyspnea, diarrhea, nausea, constipation, fatigue, and vomiting.^41,42 These findings demonstrate the need for clear communication and documentation of PROs between patients and providers as well as among members of the treatment team.

PROs are associated with clinically-important events across the disease trajectory. A meta-analysis of 21 studies reported that PROs including symptoms and quality of life were significantly associated with radiographic tumor response to chemotherapy, radiotherapy, and/or targeted therapy.⁴³ Regarding progression, Denis and colleagues have published a series of studies demonstrating that patient-reported symptoms of lung cancer (e.g., fever, cough) can identify cancer progression early (see Figure 2).^44–46 A randomized trial showed that screening for these symptoms was associated with a 7 month survival advantage over usual care, which may have been due in part to the fact that patients in the intervention group had better performance status at progression and thus were more likely to receive optimal treatment.^47,48 Lastly, several studies have shown that PROs enhance prediction of survival in myelodysplastic syndromes,⁴⁹ multiple myeloma,⁵⁰ early stage colorectal cancer,⁵¹ advanced breast cancer,⁵² metastatic castration-resistant prostate cancer,⁵³ metastatic renal cell carcinoma,⁵⁴ advanced cancers,^55,56 and a variety of tumor types.^57,58 In fact, some data suggest that PROs predict survival better than provider-rated performance status.⁵⁹

Figure 2.

A case study by Denis and colleagues⁴⁷ demonstrating changes in patient-reported outcomes (PROs) during treatment and progression of Stage IV lung adenocarcinoma. Severity of PROs is shown by light green=none, dark green=mild, yellow=moderate, and red=severe. The patient received first-line chemotherapy, then experienced two relapses, after which second-line chemotherapy and immunotherapy were administered, respectively. PROs become more severe and numerous as time goes on, with increases in severity prior to diagnosis of progression.

The Evidence Base for Clinical Monitoring of Patient-Generated Health Data

There is a growing literature base suggesting that incorporating PGHD into clinical care can improve outcomes over standard care.⁶⁰ Evidence for clinical collection of PGHD comes primarily from studies of symptom monitoring interventions of PROs, which are defined for the purposes of this review as those intended to improve communication between patients and providers regarding patients’ symptoms. Numerous randomized trials have shown that clinic-based symptom monitoring improves patient-provider communication and increases the concordance of their ratings of symptoms and quality of life.^61–69 As noted above, these benefits are consistent with additional studies suggesting that symptom management or palliative care improves survival.^{26,27,61–70} The mechanisms by which these benefits occur are currently unclear, although they may be due to better medication adherence and/or physical functioning, such that patients are able to receive more therapy.⁴⁷ Notably, improvements in outcomes have generally occurred without lengthening clinic visits.^62,64 Findings are less consistent regarding whether clinic-based symptom monitoring improves PROs, although several studies have demonstrated a beneficial effect on quality of life and/or symptomatology.^{26,64,67,71,72} Velikova and colleagues⁶⁴ found greater improvement in quality of life for patients in the intervention group when PRO data were explicitly discussed in the clinical encounter than for corresponding patients whose PROs were not explicitly discussed. A similar finding was reported by Carlson and colleagues,⁷³ in which the proportion of lung cancer patients with high distress levels were most impacted when distress screening was combined with telephone triage and resource referral, as opposed to screening alone. In contrast, there were no group differences in the percentage reporting high distress levels among breast cancer patients, who reported less baseline distress than lung cancer patients.⁷³ Similarly, McLachlan and colleagues⁷⁴ reported that, for patients in their randomized trial who reported moderate to severe depression at baseline, patients receiving clinic-based screening for unmet needs with care coordination exhibited greater reductions in their depression scores compared to patients randomized to conventional clinical care. Thus, clinic-based symptom monitoring may be most beneficial in patients with high symptomatology and when the data collected are discussed during the clinical encounter. Current literature is consistent with numerous quality of care initiatives, such as the Quality Oncology Practice Initiative (QOPI) from the American Society of Clinical Oncology (ASCO) or Electronic Clinical Quality Measures (eCQMs) from Medicare, in which EMR documentation of screening and management of PROs such as pain and distress are indicators of high-quality cancer care.^75,76

Results of randomized trials of remote symptom monitoring of PROs have been more equivocal. We define remote symptom monitoring interventions as those intended to improve communication between patients and providers regarding patients’ symptoms when patients are away from the clinic or hospital. Remote symptom monitoring often utilizes ecological momentary assessment, defined as real-time PRO reporting, that may be less subject to recall bias than retrospective questionnaires completed in clinic.⁷⁷ Although some trials have reported improvements in symptoms and quality of life,^78–80 others have not.^81–86 Mixed findings may be due to high study heterogeneity in terms of remote monitoring methodology, patient-provider communication, and the patient population. Remote monitoring methodology has included study-initiated calls from an automated telephone system,⁸⁰ patient-initiated calls to an automated telephone system,^78,81,85 study-initiated online questionnaires,^25,82 patient-initiated smartphone-based or online questionnaires,^83,87 study-initiated telephone calls from a nurse or nurse practitioner,^84,86 study-initiated calls from a research coordinator,⁸⁸ or a paper symptom diary.⁷⁹ Some studies incorporated automated alerts about severe or worsening symptoms to providers,^{78,80,81,83,85,89} whereas others encouraged patients to discuss symptoms with their providers.⁷⁹ Most studies focused on patients receiving chemotherapy or radiation^{78,81,83,84,90} or post-surgical patients,^80,86,89 although some focused instead on advanced cancer patients^79,85 or post-treatment cancer survivors.⁸² Notably, although some trials reported on process variables such as patient uptake of the intervention and provider responses to alerts,^80,81,85 no studies systematically examined important health system factors such as type of clinic or provider workflows. Nevertheless, in general, efficacious remote monitoring interventions were those that focused on patients with high unmet needs amenable to intervention, high response rates by providers to alerts, and structured provider responses that focused on symptom reduction. Taken together, these data suggest that thoughtfully-designed remote symptom monitoring interventions may improve outcomes in cancer patients. In addition, numerous studies have shown that remote symptom monitoring is feasible and acceptable to patients and providers.^{81,85,87,89–94} However, more work is clearly needed to identify characteristics of health systems, patient populations, and intervention designs to ensure that remote symptom monitoring is maximally efficacious.^95,96

Digital Phenotyping: Mobile Technology to Collect Patient-Generated Health Data

The limited use of large-scale collection of PGHD for clinical and research purposes is somewhat surprising given the abundance of opportunities to collect such data. Slow uptake may result from challenges in analyzing these data, as described below. Nevertheless, there has been a proliferation of device-based apps to track almost any aspect of health and behavior (i.e., digital phenotyping). Currently, approximately 81% of Americans own a smartphone and 17% own a smart watch.⁹⁷ Smartphone and smart watch owners tend to be younger and have a higher SES than non-owners.⁹⁷ Thus, there is concern that reliance on PGHD from these devices could exacerbate health disparities,^98,99 an issue which must be considered and addressed as PGHD collection increases. Despite this concern, commercially-available devices provide a robust opportunity to refine collection and analysis of PGHD, which in many cases has comparable reliability and validity to medical-grade biometric sensors. Device-based biometric, behavioral, real-time reporting, and performance measurement approaches are described below.

Biometrics:

The field of device-based biometric sensors is rapidly expanding. Examples include wireless scales, blood pressure cuffs, thermometers, pulse oximeters, heart rate monitors, and blood glucose meters. These typically transmit data to a smartphone app that records the readings. Smart watches are also increasingly incorporating these capabilities, such as electrocardiogram (ECG). Initial published studies have shown good accuracy in comparison to gold standard ambulatory ECG, making them useful for screening of atrial fibrillation and other cardiac issues.^100–102 Additional features in development are smartphone-based spirometry, video plethysmography, and detection of seizures.^103–105 With the exception of blood glucose monitors which face higher regulatory scrutiny, most products are regulated by the FDA as Class II medical devices, indicating that they are safe to use but not verifying their accuracy. In light of the increasing demand for device-based health monitoring, it is clear that biometric monitoring will continue to evolve in terms of accuracy and functionality.

Physical Activity:

Accelerometers have been used to quantify human physical activity in the context of research since the 1980s.¹⁰⁶ However, commercially-available wrist-worn activity trackers have only recently caught up with research-grade devices in terms of data accuracy.^107–109 The increasing ubiquity of smartwatches provides new opportunities for passive monitoring of physical activity as a remote indicator of overall health and well-being in cancer patients. Accelerometers located in smartphones can also be used to detect a variety of physical activities including walking, jogging, going up and down stairs, and sitting.^110,111 Notably, the acceptability of smartphones has been shown to be superior to that of wrist-based activity trackers.¹¹² Several observational studies have shown that higher levels of physical activity are associated with reduced risk of all-cause mortality and/or cancer-specific mortality in patients with local or regional breast cancer,^112–115 colorectal cancer,^116,117 locally-advanced non-small cell lung cancer,¹¹⁸ and other cancer types.^119,120 Higher levels of physical activity are also associated with reduced risk of several of these cancers, whereas sedentary time is associated with higher risk, independent of physical activity.¹²¹ In addition, in a meta-analysis of over one million adults from the general population, higher levels of physical activity mitigated the negative association between sitting time and all-cause mortality.¹²² Thus, remote monitoring using accelerometers could be used clinically to encourage physical activity among cancer patients.

In contrast, decreases in physical activity may indicate new-onset mental or physical health concerns among cancer patients. For example, physical slowing is a symptom of depression. Research shows that among individuals without cancer, depression is associated with subsequent sedentary behavior and activity disruption.^123–126 Sedentary behavior and smartphone use have also been found to be sensitive indicators of stress and depressive mood in non-cancer populations, such as college students.¹²⁷ Sedentary behavior, together with longer screen time on a smartphone, was found to identify symptom burden among cancer patients with an accuracy of 88%.¹²⁸ Moreover, sudden decreases in physical activity during active treatment may indicate acute toxicities such as renal insufficiency, pneumonitis, and gastritis, although more research needs to be conducted on these relationships.

Location:

Patients’ community mobility may also indicate overall mental and physical well-being. Several methods have been developed to measure location using smartphones, including global positioning system (GPS), Wi-Fi, and Google Location History.^129–132 Location may be a proxy for a diversity of daily activities outside the home that indicate good quality of life, including social, leisure, and work activities as well as activities of daily living (e.g., shopping). Regarding mental health, in a study of adults without cancer, depression was detected with an accuracy of 87% using GPS measurements of movement, mobility between favorite locations, and movement independent of location.¹³³ Another study found that GPS monitoring in combination with location data from the search-and-discovery app Foursquare identified depression and anxiety with accuracy of 88%, better than either alone.¹³⁴ Regarding physical health, better physical quality of life and less pain have been associated with number of trips outside the home in patients recovering from stroke and spinal surgery, respectively.^135,136 GPS can also be used to track outdoor physical activity such as walking, which has been found to occur less often in non-small cell lung cancer patients as compared to similar individuals without cancer.¹³⁷ Outdoor walking capacity using GPS has found to be strongly associated with clinic-based physical performance measures such as the six-minute walk test.¹³⁸ Thus, there is significant potential to use location data as a way to remotely monitor cancer patients’ quality of life.

Diet:

Dietary habits have been notoriously difficult to accurately capture due to the use of long retrospective questionnaires that can be susceptible to recall bias and social desirability bias.¹³⁹ Due to participant burden, dietary habits are typically measured relatively infrequently in epidemiologic studies, which may impact study findings. Web- and app-based dietary measures offer opportunities for improved measurement.^140,141 Participant burden can be reduced through use of branching logic for question presentation. Additionally, accuracy of recall may be improved through use of photos of portion sizes that would be difficult to present on paper-and-pencil-based measures.^142–144 Although not fully developed yet, the use of a smartphone camera together with machine learning algorithms to identify foods and portions may be a less-burdensome way to collect dietary information in the future.^145,146 Nevertheless, studies of device-based measurement of dietary habits tend to suffer from high rates of attrition, which itself may bias results. Assessment of diet is clinically important, however. A diet that is low in saturated- and trans-fat and high in fiber, vegetables, fruits, and other nutrients may reduce risk of all-cause and cancer-specific mortality after diagnosis of breast or colorectal cancer,^147–150 although evidence is limited in patients with other cancer types.^151,152 Interestingly, patients who improved their diet after colorectal cancer diagnosis also demonstrated lower cancer-specific mortality,¹⁵⁰ suggesting that Web- and app-based dietary measurement could be used in the context of clinical intervention to improve cancer outcomes.

Sleep:

There is a large body of literature to support the reliability and validity of sleep assessment using research-grade, wrist-worn accelerometers.¹⁵³ Outputs include amount of time awake after sleep onset (WASO), sleep efficiency (i.e., percentage of time in bed spent sleeping), and total sleep time. Traditionally, assessing sleep stages (e.g., REM sleep) required polysomnography in sleep laboratories. Over time, the validity of consumer-grade wearable devices has improved, and some devices now include a heart-rate sensor which enables detection of some sleep stages. Some of the more recent devices also demonstrate similar agreement with gold-standard polysomnography and research-grade devices.^154–157 However, one drawback of wrist-worn devices is that they will likely need to be recharged during the assessment period. Project Baseline has instead used a piezoelectric sensor placed under the mattress to detect sleep, which has demonstrated high validity and does not require recharging.¹⁵⁸ A less accurate, but very low burden, detection method is to use smartphone screen interaction to approximate bedtimes and waking times.¹⁵⁹ Smartphones have also shown promise for evaluation of perceived sleep quality using patient reports as well as detection of sleep-disordered breathing using built-in microphones.^160–163 In addition, the MyHeart Counts study from Apple uses smartphones to detect ambient light.¹⁶⁴ Nevertheless, until battery life and sleep sensor technology improve in commercially-available wearables and smartphones, alternate measures of sleep may need to be utilized.

Real-Time Reporting:

Mobile devices provide an outstanding opportunity to capture real-time patient-reported data. There is a large literature base focused on ecological momentary assessment (EMA) of symptoms such as pain, fatigue, and depression in real-time. Although burdensome for patients relative to passive monitoring of behaviors, EMA can provide important information regarding daily and intra-day variability in symptoms without recall bias.¹⁶⁵ For example, using real-time, validated, smartphone-based “brain games,” Small and colleagues¹⁶⁶ demonstrated significant intraday cognitive variability in breast cancer patients. Cognitive variability may be a more sensitive indicator of cancer-related cognitive impairment than office-based neuropsychological assessment, as it precedes cognitive decline in normal aging.^167,168 Smartphone-based cameras are increasingly used to remotely assess dermatologic conditions or surgical complications,^169–172 and have the potential to identify complications early, providing better, lower-cost care. Real-time reporting can facilitate just-in-time or micro-interventions, cognitive-behavioral interventions that can be delivered via smartphone or smartwatch to deliver interventions personalized to an individual’s current circumstances or environment.^173,174 Examples include reminders to exercise during periods of inactivity or smoking cessation texts to suggest coping strategies during reported nicotine cravings. Lastly, an innovative clinical trial has used smartphone-based real-time symptom reporting to manage dose-limiting toxicities of cediranib and olaparib for ovarian cancer, a regimen that can cause rapid-onset hypertension and diarrhea.⁷ There is untapped potential for remote symptom monitoring in clinical trials to keep patients on drug longer, with potential survival benefits.

Functional Status:

A sizable body of research has demonstrated that smartphones can reliably detect gait, physical performance, range of motion, and falls. A smartphone placed in a pants pocket can detect stride times, a clinically-meaningful metric of locomotor control, with comparable accuracy to gold-standard clinic-based instrumentation.¹⁷⁵ Interestingly, measurement of gait parameters is equally reliable and valid regardless of whether the smartphone is worn on the body, on a belt, or in a pocket or purse.^176,177 Two- and six-minute walk tests can be administered remotely via smartphone with comparable results to clinic-based tests.^178–180 The same is true for the Timed Up and Go (TUG) Test, in which a patient is asked to stand up from a chair without using their arms, walk a set distance, return, and sit down again.^181,182 Range of motion in the wrist¹⁸³ and ankle¹⁸⁴ can also be captured reliably via smartphone. Additionally, falls can be distinguished from daily actions with an accuracy of 90% using a smartphone-based accelerometer and gyroscope.^185,186 Similar to gait detection, the accuracy of fall detection does not appear to differ based on the placement of the smartphone on a belt or in a pocket.¹⁸⁷ Use of a smart watch together with a smartphone can significantly decrease the rate of false positives when detecting falls.¹⁸⁸ Remote monitoring of functional status may be particularly useful for telemedicine visits and in geriatric populations.

Big Data: Untapped Potential for Patient-Generated Health Data

Digital phenotyping can yield very large datasets amenable to big data analytics. The term “big data” is defined by Google as “extremely large data sets that may be analyzed computationally to reveal patterns, trends, and associations, especially relating to human behavior and interactions.”¹⁸⁹ Initial forays into big data in the context of healthcare have been limited primarily to analysis of clinical information, payment and billing data, genomics, and biomarkers. Little has been done specifically in PGHD. Historically, a primary concern regarding sharing of health data has been patient privacy. The U.S. Healthcare Insurance Portability and Accountability Act (HIPAA) of 1996 and later laws focused on regulating electronic storage, transmission, disclosure, and reporting of security breaches of identifiable health information, as well as enforcement mechanisms for healthcare systems that did not comply.¹⁹⁰ The U.S. Health Information Technology for Economic and Clinical Health (HITECH) Act of 2009 developed a system of incentives and penalties for healthcare systems to implement electronic medical records (EMR) systems.¹⁹¹ The widespread implementation of EMRs has enabled large-scale pooling of deidentified data from patients outside the context of a clinical trial. Concurrently, it has become increasingly clear that very large datasets would be required to detect and understand the complex genetic and biological mechanisms of cancer.

At the same time, a consensus emerged that the field of medicine had an ethical obligation not just to protect patients’ privacy but also to aggregate and learn from their data. This position was articulated in 2006 in a publication from the IOM entitled “The Learning Healthcare System.”¹⁹² Alarmed at the rising costs of healthcare, the IOM recognized an urgent need for data to optimize the effectiveness and value of medical decisions. The existing framework of randomized clinical trials and regulatory approvals was too slow and costly to provide evidence for more than a sliver of decisions made by healthcare providers on a daily basis. As a result, the IOM proposed using real-world evidence to fill in knowledge gaps in the fast-changing healthcare landscape. The IOM recognized that progress towards this goal would rely on “the emergence of linked clinical information systems that might allow information about safety and effectiveness to emerge naturally in the course of care.”¹⁹² A follow-up IOM conference¹⁹³ set forth a rapid-learning systems model specifically for oncology. The 2009 conference provided a more detailed vision of an iterative process in which clinical data are systematically collected and analyzed, with the resulting insights implemented into clinical care. Changes in outcomes are then measured and form the basis for new hypotheses, analyses, and adjustments to care.¹⁹³ Although rapid-learning systems were described with a focus on clinical data, it is not hard to envision incorporation of PGHD as well. The vision of a rapid learning system relies heavily on advancements in large-scale data aggregation, harmonization, and analysis that are only now starting to be realized.

Recent large-scale data collection initiatives can be seen across government, advocacy, and academic sectors. Regarding government, in 2015 President Obama announced a new Precision Medicine Initiative, one focus of which was to create infrastructure for open data sharing. The 21^st Century Cures Act in 2016 provided $1.8 billion in funding to support the Cancer Moonshot Initiative to build a national cancer data infrastructure, establish networks for patients to directly contribute tumor profile and PRO data, and analyze and biospecimens from past clinical trials to predict future patient outcomes. The National Institutes of Health (NIH) is also directing the All of Us research program, an ambitious cohort study of a million or more individuals of diverse backgrounds living in the United States.¹⁹⁴ Participants will be followed for ten years; data collected include patient-reported surveys, clinical data from the EMR, and biospecimens.¹⁹⁴ The NIH also maintains a wide variety of additional federal health databases of clinical, genomic, imaging and epidemiological data from federally-funded research projects.¹⁹⁵ These federal initiatives are consistent with those launched by advocacy groups. One example is ASCO’s CancerLinQ, a platform for sharing and analyzing oncology data from healthcare IT systems.¹⁹⁶ Currently consisting of over a million records from 2,000 oncology care providers, the goals of CancerLinQ are to improve quality of care and facilitate research.¹⁹⁷ Other advocacy groups, such as the GO2 Foundation for Lung Cancer, have patient registries in which patients can share data with researchers.¹⁹⁸ Indeed, cancer patients express willingness to provide PGHD to cancer registries if data are kept confidential.¹⁹⁹ Academic institutions are also undertaking large data collection efforts. An example is the Oncology Research Information Exchange Network (ORIEN), a partnership among 19 academic cancer centers to collect and aggregate deidentified molecular, clinical, and epidemiological data.²⁰⁰

The private sector has launched parallel initiatives. For example, Flatiron Health created a computing platform to provide EMR, billing, analytics, and clinical trial screening capabilities for community oncology practices. Curated clinical data from the platform flow into a data warehouse, which can be mined to monitor quality of care, examine patterns of real-world treatment utilization, and create in silico control groups for clinical trials. Another example from the private sector is Verily’s Project Baseline Health Study, a four-year longitudinal study of 10,000 individuals living in the United States. Data to be collected include biometric data (e.g., electrocardiogram data, electrodermal data, heart rate, sleep quality) from Google Watch, genetic information, blood samples, clinical data from annual in-person study visits with healthcare providers, and PROs. This project has recognized the vast analytic possibilities engendered by mobile and wearable devices. Google has also partnered with healthcare providers such as Ascension, an organization consisting of 2,600 hospitals, doctors’ offices, and other facilities, to aggregate patients’ health histories. The goal of this initiative, called Project Nightingale, is to design new artificial intelligence algorithms to support patient care. Similar initiatives are underway at Apple, Microsoft, Amazon, and other companies. An example is the Apple Health Records application programming interface (API), under development as part of HealthKit, an app that aggregates and presents health information from iPhone, Apple Watch, and third-party apps. Health Records can link to the EMR or patient portal and display information on health conditions, lab results, immunizations, medication, procedures, and other data from participating medical providers. Apple also allows users to share health data with researchers. More broadly, APIs allow for data from devices and wearable sensors to be uploaded to the EMR or data warehouses with the patient’s permission. Thus, preliminary infrastructure resources needed for large-scale PGHD collection have been established, from which additional APIs can be developed. In the meantime, most large aggregated datasets are typically comprised of clinical and molecular data with little to no PGHD.

Data Mining, Natural Language Processing, and Artificial Intelligence Using Big Data: A Promising Approach for Analysis of PGHD

The term artificial intelligence was first coined in 1956 as “the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.”²⁰¹ Concurrently, there was recognition that algorithmic decision-making might actually surpass the accuracy and reliability of human judgment, an assertion famously proposed by the psychologist Paul Meehl in 1954²⁰² which has been supported by an impressive array of data.^203–205 Although AI was dismissed by the New England Journal of Medicine in 1987²⁰⁶ because “the field of medicine is so broad and complex that it is difficult, if not impossible, to capture the relevant information in rules,” it experienced a resurgence in the early 2010’s. This development was due to the convergence of three major trends: 1) improvements in computing power; 2) development of new AI algorithms; and, 3) increases in the number and quality of big data sets for training.²⁰⁷ The first two trends allowed scientists to construct and train dramatically larger and more complex “deep” neural networks. The size and quality of training data, meanwhile, allowed these networks to achieve high accuracy at several narrow-domain tasks. One early application was classification of images,²⁰⁸ followed shortly by labeling and outlining (semantic segmentation) of multiple objects within images.²⁰⁸ These developments were transferred and applied by biomedical scientists to classification of medical images and detection of skin cancer.^209,210 There has also been increasing application of machine learning to PGHD. For example, two recent studies have demonstrated that PGHD in the form of internet search logs of symptoms facilitate early identification of cases of pancreatic cancer and lung cancer.^211,212 Machine learning has also been applied to PROs to predict the course of multiple sclerosis^213,214 and recovery after hip and knee replacement.²¹⁵

Artificial intelligence is rapidly evolving. For example, in late 2018 scientists at Google published a powerful new algorithm for natural language processing: Bidirectional Encoder Representations from Transformers (BERT),²¹⁶ a powerful new algorithm for natural language processing (NLP) developed by Google. BERT set new performance records across nearly all NLP benchmarks, and for the first time, achieved human level performance on several of these. BERT and related technologies are being applied to unstructured biomedical^217,218 and clinical texts.^219,220 Translation of these new technologies into the clinical domain is still in the early stages but progress is accelerating. In 2018 the FDA developed a “fast-track” approval process for AI-based medical technologies.²²¹ By late 2019 the FDA had cleared 26 AI-based tools for marketing and use in the US.²²² Although there are no commercial AI tools to our knowledge focused on PGHD in the context of cancer, we expect such tools to appear shortly. Sufficient computing power and algorithms are widely available. These should allow scientists to identify and extract new information from PROs, wearable sensors, and EMR notes.

Challenges and Opportunities for Clinical Integration of Patient-Generated Health Data

Although advancements in technology have the potential to revolutionize medicine through collection and analysis of PGHD, clinical integration of PGHD has lagged behind. Challenges to clinical integration include data linkage and scaling across healthcare systems, provider engagement, and actionability. A significant barrier to data linkage and scaling across healthcare systems is the inflexibility of most EMR platforms. While commonly-used EMRs in oncology such as EPIC are starting to incorporate PRO measures, uptake of both PROs and biometric data has been limited by usability issues of the EMR platforms themselves. As the IOM noted in 2013,²²³ “Originally designed for billing and coding purposes, health IT systems have not been integrated efficiently into clinical care, do not facilitate the coordination of care, and the need to customize local systems has created a situation where health IT systems cannot communicate with each other… Many of these systems are inflexible and thus are unable to adapt to the changing needs of a modern health care system.” When collected as part of clinical care, PGHD may be buried in the clinical record, with poor visibility and interpretability. For example, Rotenstein and colleagues²²⁴ reported barriers to use of PRO data in a large radiation oncology clinic including difficulty accessing the data in the EMR (60%), difficulty bringing the data into the treatment note (48%), too much data (38%), and difficulty interpreting data (26%).²²⁴ Notably, these barriers also apply to remotely-collected biometric data and could be addressed through better EMR integration and visualization. In recognition of such barriers to clinical integration, the National Cancer Institute’s Center for Strategic Scientific Initiatives included in its 2020 provocative questions the need to identify new ways to integrate PGHD into the EMR. The development of the Fast Healthcare Interoperability Resources (FHIR) data standard and APIs provides opportunities to flexibly create software that securely pulls discrete data from the EMR into third-party software. FHIR could be used to improve presentation and visualization of PGHD in the form of both PROs and remotely-collected biometric data, although EMR companies must enable FHIR capability. Recent trends in information technology (IT) define FHIR as an approach suitable for citizen developers, in which applications are developed by non-experts with the support and approval of cooperating IT experts.²²⁵ Driving this trend is the development of low-code/no-code solutions, which enable users with little to no programming background to create and interact with the applications they need. Numerous potential benefits include rapid development, no need to wait for developers, and solutions that solve the problem at hand. There are also potential drawbacks, including security, scalability, and limited use for complex application. These IT trends may improve the access and impact of the collected data for patients, researchers, and physicians. Developers of such software should draw heavily from research in the fields of user-focused design and human-computer interaction to create intuitive data visualizations that focus on important clinical information and allow providers to identify patterns of data that provide important clinical insights. Although some initial research has been conducted on visualization of PGHD,^226–229 transdisciplinary collaborations including citizen developers, graphic designers, cognitive scientists, bioinformaticians, biostatisticians, data scientists, and computer scientists will be needed to create the frictionless user experience necessary for optimal uptake.

Despite challenges with EMR integration, clinical collection of PROs in oncology results in high provider satisfaction.^{224,230–232} High rates of satisfaction have been reported despite initial provider concerns. These initial concerns commonly focus on the time required to collect and address PROs in clinic, lack of training regarding interpretation and management of some PROs, patient burden, and liability issues if PROs are overlooked.^233–235 Thus, an important aspect of successful clinical implementation of PROs is addressing provider concerns, including appropriate provider training and clinical decision support.^{13,14,19,236,237} Concerns may also be ameliorated in part by communicating observed benefits of PROs reported by other providers. PROs can contribute to shared decision-making and strengthen rapport between patients and providers by facilitating better communication.²³² Interestingly, PROs may also reduce provider burnout by contributing to more efficient clinical workflow.²³² Because providers arrive to the clinical encounter with information regarding patient concerns, they can spend more time addressing these concerns in a meaningful way rather than running through checklists of questions.²³² PRO data may also be less vulnerable than traditional clinical assessments to “white coat syndrome,” in which patients are less likely to report symptoms during an interview than on paper.²³⁸ Following clinical rollout of PROs in a radiation oncology setting, over half of the 53 providers surveyed reported no or rare impact of PROs on their ability to see patients on time and a majority would recommend PROs to a colleague at another institution.²²⁴

Several steps can be taken to address actionability concerns as well. Perhaps most importantly, collection and use of PGHD to drive clinical decisions should be incorporated into medical boards and continuing education. Training should be extended to physicians, advanced practice professionals, nurses, and medical assistants. Training should include both management of PROs and use of PGHD to support shared decision-making. For example, Thomas Jefferson University has a graduate certificate program in digital health. Organizations such as ASCO and the National Comprehensive Cancer Network (NCCN) provide detailed, evidence-based guidelines for management of PROs including nausea and vomiting, fatigue, distress, and general supportive care. Some PRO data collection apps, such as Carevive, use algorithms based on these guidelines to suggest self-management strategies.²³⁹ Measure-specific algorithms for identifying severe symptoms and clinically-significant worsening of symptoms have been well-articulated.^240,241 The necessary elements are in place to determine actionability of PROs. Better provider education and well-integrated clinical decision support remain to be addressed, however. Payer reimbursement for symptom management, particularly in the realm of telehealth and digital therapeutics, is also needed. Efforts to address these issues should include the perspectives of multiple stakeholders, including patients, families, providers, payers, EMR companies, and IT developers.²⁴²

Summary and Future Directions

The past decade has seen unprecedented progress in the war on cancer in both therapeutic and technological domains. We now have the opportunity to integrate these accomplishments by creating new, data-driven approaches using PGHD to identify and intervene early on clinically-significant events such as toxicity and cancer progression, which may in turn reduce emergency room visits and hospitalizations. At the same time, technology allows closer contact with patients outside of the clinical encounter to facilitate healthy lifestyle behaviors, symptom management, and medication adherence. There are still many technological, analytic, and workflow challenges that need to be overcome to incorporate PGHD into routine research and cancer care, however.^{13,17,242–244} The evidence base to support use of biometric data is sparse relative to that for PROs. In addition, PGHD must be implemented in a way that prevents exacerbation of health disparities in oncology care. Nevertheless, as with treatment discoveries, although each intervention is incremental, together PHGD-based interventions can dramatically improve both quality and quantity of life.