Traditional clinical data—laboratory values, imaging results, and vital signs recorded during visits—are generated through standardized protocols controlled by health care providers. Measurement timing follows clinical indication, professional judgment, or research schedules, and continuous professional oversight typically limits the influence of patient behavior on data availability.

Patient-generated telemedicine data invert this relationship. The patient controls when, whether, and how data are recorded [21,23]. Device use patterns, self-report compliance, and engagement with monitoring platforms all depend on patient decisions embedded in daily life contexts. This shift introduces behavioral factors as primary determinants of data availability—and, consequently, as potential drivers of missingness mechanisms.

It is important to distinguish 2 sources of data incompleteness. Adherence-driven missingness arises from patient behavior: decisions not to measure, forgetting, disengagement, or intentional avoidance. System-driven missingness arises from technical factors: device malfunction, battery depletion, connectivity failures, or sensor drift. While both produce gaps in the observed record, their implications differ: adherence-driven missingness more directly invites behavioral or clinical interpretation, whereas system-driven missingness first requires technical quality control. In practice, the 2 often co-occur and can be difficult to disentangle—a patient who removes a wearable device exhibits both technical nonwear and behavioral choice.

Research consistently documents declining engagement with digital health interventions over time [9,10]. Large-scale analyses involving over 100,000 participants confirm that early engagement patterns strongly predict long-term retention [9], and this decline follows trajectories influenced by patient characteristics, intervention design, and contextual factors [24-26]. Critically, adherence patterns are not merely predictors of data availability—they are associated with health outcomes across chronic pain [27], rheumatoid arthritis [28], smoking cessation [29], mental health [30], and cardiac rehabilitation [31]. App engagement also moderates intervention effectiveness [32].

This evidence points to a structural feature of patient-generated telemedicine data: the data we observe are systematically filtered through an adherence process that may itself carry clinical meaning. Patients who contribute the most data are often those who are most engaged; those whose data are most sparse may be the patients about whom we most need information.

A substantial strand of applied digital health and wearable-sensing research approaches missing observations primarily as a reconstruction or preprocessing problem—asking how accurately missing values can be estimated, replaced, or made usable for downstream modeling [33,34]. We cite such studies not as examples of poor methodology but as illustrations of a legitimate and influential analytical orientation: missingness is treated primarily as a target for reconstruction or preprocessing. This concern is not merely theoretical. Goldberg et al [11] found that differential attrition was common in mHealth randomized trials, yet sensitivity analyses for data MNAR were not reported in the reviewed studies. However, reconstruction is not the same as interpretation. Methods designed to estimate missing values ask what the missing measurement might have been, not why it is missing. A missing entry may reflect forgetfulness, technical failure, symptom worsening, treatment burden, or deliberate disengagement—causes with different clinical and analytical implications even when they produce the same empty cell. If missingness is adherence driven, imputing values without first characterizing the missingness pattern may erase information about patient behavior, differential engagement, or clinical vulnerability. The ADSF complements reconstruction-oriented approaches by shifting the first analytical question from “How should we fill the gaps?” to “What generated these gaps, and should they be interpreted, modeled, reported, or only then reconstructed?”

The data characteristics and practice gap described in the previous section manifest in 4 interconnected challenges for telemedicine data science. We formulate these as principles that, taken together, constitute the conceptual foundation of the ADSF.

Each principle is grounded in established findings: principle 1 draws on documented engagement decline patterns in mHealth [9,10] and the resulting data irregularity [23,35]; principle 2 builds on the foundational missingness taxonomy [14,15] and evidence of differential attrition in mHealth trials [11]; principle 3 extends well-documented population heterogeneity in treatment response [36,37] to its interaction with adherence-driven data availability; and principle 4 responds to concerns about model-data mismatch in health care AI [38,39].

Patient-generated data exhibit sparsity across multiple dimensions [23,35]: irregular sampling intervals, missing scheduled measurements, multi-resolution heterogeneity across data streams, and fragmented trajectories with variable-length gaps. These are not incidental data quality issues; they are direct consequences of placing data generation in patients’ hands.

Sampling irregularity degrades the performance of many standard analytical approaches [36]. While several architectures have been designed to handle missing values and irregular sampling directly—including the gated recurrent unit with decay (GRU-D) model [37], convolutional approaches for incomplete data [38], and attention-based methods [39,40]—the challenge in patient-generated telemedicine data is that missingness rates often far exceed those in the datasets on which these architectures were developed, and the missingness process itself may be informative rather than ignorable.

Patients differ substantially in adherence patterns, disease trajectories, and treatment responses. This heterogeneity interacts with adherence-driven sparsity to create compounding analytical challenges. Population-average effects frequently fail to predict individual responses [43,44], and person-specific analyses reveal that relationships observed at the population level may not hold, or may even reverse, at the individual level [44].

Heterogeneity amplifies the sparsity problem in 2 ways. First, naive pooling of sparse observations may obscure person-specific relationships, and individual-level interpretation may require methods that preserve within-person temporal structure. Second, different patients may exhibit different missingness mechanisms: some disengage when symptomatic, others disengage when doing well. This heterogeneous missingness violates uniform assumptions underlying standard imputation approaches [11,45] and means that a single missingness-handling strategy may be inappropriate for the entire sample.

The practical consequence is differential representation: high-adherence patients contribute dense trajectories that dominate analyses, whereas low-adherence patients contribute sparse fragments that may be excluded or downweighted. Models built on such data risk reflecting only the experience of engaged patients and failing to generalize to the broader patient population.

Contemporary health care AI has pursued increasingly complex architectures—deep learning, transformers, and multimodal fusion—that typically require large, high-quality datasets [46]. Patient-generated telemedicine data may instead involve small samples, high sparsity, heterogeneous populations, and potentially informative missingness [11,13,47].

This mismatch has practical consequences. Models may achieve high apparent performance by exploiting informative missingness patterns rather than true clinical signals, undermining transportability when missingness patterns change across settings [48]. The interpretability required for clinical adoption may conflict with the opacity of sophisticated architectures [49]. In many sparse or small-sample clinical settings, simpler interpretable models may be competitive with more complex deep learning approaches [47,50].

The alignment calls for honest assessment of actual data characteristics. A central concern is the uncritical transfer of methods from domains with dense, regularly sampled data (eg, imaging and financial time series) into patient-generated telemedicine research, where observations are sparse, irregular, and shaped by adherence—methodological sophistication alone is insufficient without alignment with the data generation process.

The ADSF workflow proceeds through 3 stages. Figure 1 provides an overview.

Stage 1 (identifying the data generation context) determines the extent to which data availability depends on patient behavior and distinguishes adherence-driven gaps (the patient did not log in, removed the device, or disengaged) from system-driven gaps (server downtime, battery failure, or connectivity loss) using device logs, time stamps, or platform metadata where available.

Stage 2 (characterize the data) systematically assesses data properties using the reporting dimensions described in the Data Characterization: A Reporting Framework section, producing a structured profile that makes adherence-driven sparsity visible and reportable.

Stage 3 (align analytical reasoning) uses the data profile to guide methodological choices following the principles in the Four Methodological Principles section, ensuring that analytical decisions are defensible given the characterized data properties.

The following methodological principles operationalize the 4 conceptual principles described in the Adherence-Driven Sparsity: 4 Principles section: principle A follows from recognizing missingness as information (principle 2); principle B follows from mechanism plausibility (principle 2); principle C follows from sparsity, irregularity, and heterogeneity (principles 1 and 3); and principle D follows from model-data alignment (principle 4). These principles illustrate compatible analytical reasoning rather than prescribing specific algorithms.

The following is a hypothetical illustrative example with plausible but simplified values intended to demonstrate ADSF reasoning rather than report empirical findings. Consider a 12-week smartphone-based depression monitoring study with 200 participants completing daily mood logs, with baseline Patient Health Questionnaire–9 (PHQ-9) scores available.

Data characterization is as follows. Data source is active daily mood logging via smartphone app, indicating high dependence on patient behavior. Missingness amount is 58% of expected entries, with per-patient completeness ranging from 4% to 100% (mean 42%). Temporal pattern is declining completion over weeks 6 to 12 in most participants, with a subset showing episodic re-engagement (systematic decay with nonmonotone elements). In terms of mechanism plausibility, higher baseline PHQ-9 scores are associated with higher dropout (odds ratio 1.8; P<.01). This pattern is compatible with MAR because baseline PHQ-9 is an observed variable. However, missing entries also cluster before clinical deterioration events (hospitalizations and crisis contacts), raising concerns that missingness may additionally depend on the patient’s evolving but unobserved symptom state—making MNAR plausible. In this scenario, clinical deterioration and adherence are linked rather than independent: worsening symptoms may directly reduce the patient’s capacity or motivation to engage with monitoring, making the resulting missingness adherence driven rather than random. The systematic clustering of gaps before clinical events is inconsistent with MCAR, which requires that missingness probability is independent of both observed and unobserved values. This does not prove MNAR but justifies sensitivity analyses under MNAR assumptions. Device logs show that 12% of gaps coincide with server downtime or app crashes (system driven). In terms of patient heterogeneity and technical context, the intraclass correlation coefficient for mood trajectories is 0.62 (substantial between-person variability); server logs confirm that 12% of gaps are technical.

Analytical reasoning (following principles A-D) involves including gap duration, completion slope, and days since the last entry as features (principle A); fitting a joint longitudinal survival model with sensitivity analyses under MAR vs MNAR (principle B); using the Gaussian process or mixed-effects models as primary analysis for irregular observations (principle C); and, with N=200, prioritizing interpretable probabilistic approaches over complex deep learning architectures (principle D).

This example illustrates the intended role of the ADSF: not to select a method automatically but to make visible the assumptions and behavioral interpretations that should precede method selection.

To illustrate the difference that ADSF-guided reasoning can make, consider an alternative analysis of the same data that follows standard practice without systematic missingness assessment. A researcher might apply multiple imputation under a default MAR assumption without separating the 12% system-driven gaps from adherence-driven gaps, without assessing whether the baseline severity-dropout association fully accounts for the missingness pattern, and without preserving gap duration or completion trajectories as analytical features. Under this approach, the association between symptom worsening and predeterioration disengagement would be treated as missing information to be reconstructed rather than as a potential behavioral signal to be investigated. The ADSF does not guarantee different conclusions, but it ensures that such decisions are made explicitly rather than by default.

The ADSF has practical implications across the research life cycle.

The ADSF draws on and integrates several established lines of research. The missing data literature [14,15] provides the foundational taxonomy of missingness mechanisms. The concept of structured missingness [22] recognizes that missingness patterns in biomedical data can carry systematic, leverageable structure. The mHealth engagement literature [12,13] documents the prevalence and clinical correlates of adherence-driven data loss. Finally, the behavioral science perspective [16,18] explains why engagement is not merely a technical behavior but a psychologically meaningful process.

Among these lines of research, Scherer et al [21] demonstrated that engagement and dropout can be jointly modeled using shared random effects, providing important empirical evidence that missingness and engagement are statistically linked in mHealth data. The ADSF builds on this insight but operates at a different level: rather than proposing a specific statistical model, it provides a preanalysis workflow for determining which analytical approach—including but not limited to joint modeling—is appropriate given the characterized data properties.

The contribution of the ADSF is not to any of these individual areas but to their integration in a specific applied context. By connecting behavioral understanding of why patient-generated data are incomplete with methodological reasoning about how to analyze them, the framework addresses a gap that exists between disciplines rather than within them. The relevant knowledge exists across these disciplines; what has been lacking is a structured workflow that integrates behavioral understanding of why data are incomplete with methodological reasoning about how to analyze them in the specific context of patient-generated telemedicine data. The ADSF provides this integration through a shared vocabulary and a systematic preanalysis process.

Several limitations warrant acknowledgment. First, the ADSF is a conceptual and interpretive framework that has not been empirically validated through comparative studies. Future research should evaluate whether ADSF-guided reasoning leads to more defensible conclusions than standard approaches across diverse telemedicine datasets.

Empirical demonstration of the ADSF using real-world telemedicine datasets is an important direction for future research. Such work could compare analytical conclusions reached with and without systematic adherence-driven sparsity assessment, providing evidence for the framework’s practical value.

Second, the framework does not provide software implementations, and effective application requires both statistical expertise and clinical domain knowledge. Assessing missingness mechanisms, implementing sensitivity analyses, and interpreting ambiguous patterns are nontrivial tasks that may limit adoption in resource-constrained settings.

Third, not all missingness in patient-generated data is behaviorally meaningful. The framework may encourage overinterpretation of gaps that are random or technical; the reporting dimensions in Table 1—particularly the technical context dimension—guard against this, but the risk remains. Relatedly, the framework’s applicability may vary across clinical domains, measurement modalities, and patient populations, and domain-specific adaptations are needed.

Fourth, interpreting missingness as a potential clinical signal has ethical and operational implications. False alarms based on engagement decline may increase clinician workload or patient burden, whereas missed signals may create safety risks. Any clinical use of adherence-based alerts requires prospective validation, appropriate safeguards, and careful consideration of patient autonomy and privacy.