We searched PubMed, Embase, EBSCO, Ovid, Scopus, Web of Science, and Springer Link. Search terms included terms describing the concept (health portrait) in combination with population (NCDs) and context (big data). We also included related terms such as “User-centered Design” and “Telemedicine” as search terms because they may be related to the concept or context. The search was restricted to articles published in English or Chinese from January 2014 to July 2024. Furthermore, we scanned reference lists of included publications and published reviews for additional articles. The full search queries can be found in Multimedia Appendix 2.

Population is limited to patients diagnosed with NCDs defined by the World Health Organization [45]. Moreover, we also included studies on a broader range of NCDs where relevant, like obesity, mental disorders, and geriatric syndromes, to ensure a comprehensive review of health portrait types and the methodologies used [46,47].

Concept (health portrait) is defined in this study as the models or programs to stratify or classify patients under a certain characteristic or condition based on characteristic information and process data labels related to health status [15], for example, specific social determinants, health risk scores, and lifestyle behaviors.

Context (big data) is defined as the extent to which the study’s modeling process aligns with the “3Vs” framework—volume, velocity, and variety. Specifically, a study is considered to meet the criteria if it satisfies all of the following: (1) the inclusion of unstructured data; (2) the use of data sources from digital interactive platforms; and (3) the incorporation of patient data covering natural, domain, and specific attributes.

The types of publications included observational studies and interventional studies. Besides, we excluded studies that used only traditional data collection methods (interviews or scales) for data collection. Studies that used public databases, EHR systems, digital application platforms, and wearables and sensors for data collection are included. Detailed explanations about the concept (health portrait), population (NCDs), and context (big data) can be found in Multimedia Appendix 3. The study screening manual is shown in Table 1.

To increase the consistency of study screening among reviewers, reviewer 1 piloted the study screening manual for database search and study selection based on title and abstract information available in the databases. Subsequently, reviewer 2 independently cross-checked the study selection of all articles identified in the database search. Both reviewers then discussed results and amended the screening manual before the data charting step. Subsequently, both reviewers 1 and 2 independently piloted the study screening manual for evaluating the eligibility of 10% of all identified full-text reports using a computer-generated random sequence, along with complete data charting for the included articles. Both reviewers then discussed the results and amended the screening manual. Finally, reviewers 1 and 2 independently completed an assessment of the remaining full-text reports for eligibility, along with data charting for all included reports.

This review included both observational studies and interventional studies. To reconcile the scope review’s mandate for breadth with methodological accountability, we used a dual-tiered screening framework according to some previous research [48,49]: (1) prioritizing journals to leverage rigorous peer-review standards and (2) inclusion of methodologically innovative studies despite their limited overall quality, following thorough internal team deliberations to ensure balanced inclusion. While formal quality appraisal tools were intentionally omitted to preserve conceptual mapping flexibility—an accepted scoping review limitation—we embedded critical safeguards: systematic exclusion of studies with high bias risks (randomization, blinding, and allocation concealment) and study type-specific quality screening aligned with Multimedia Appendix 4. This included external validation thresholds for clinical predictive models, health recommender systems, and interventional studies.

Reviewer 1 extracted basic information from the included articles provided by the first author: published year, country, study design, participants, modeling approaches, external validation status, sample size, and data usage. We group studies by the scenes, participants, objectives, and modeling approaches of the health portraits, using Python-extracted keywords in the titles and abstracts, and reviewers conduct secondary integration of keywords. An expert panel and the research team determined the 3V framework based on the concept of big data and the comprehensive capability assessment (Multimedia Appendix 5). All findings are synthesized to identify existing status and knowledge gaps for future research. Key results are summarized using the word cloud analysis to identify different types of health portraits, the heat map to identify the existing status of health portraits that meet the requirements of “3V,” as well as the spider diagrams to perform the comprehensive capability of included records meeting the requirements of both the external validation and “3V.” Any disagreements on study selection and data charting during pilot testing were resolved by consensus, or otherwise with a tiebreaker by reviewer 3 if needed.

The scatter plot (Figure 2) highlights key terms such as “management,” “medical,” “clinical,” “community,” “life,” “home,” and “trajectory,” reflecting the diverse application of health portraits in inpatient and outpatient settings for managing NCDs. Participants in these systems include medical professionals, patients, caregivers, and virtual health coaches, as indicated by terms like “patients,” “experts,” “caregiver,” “coach,” “chatbot,” “nurse,” and “self.” This underscores a collaborative approach spanning hospital care, community support, and patient self-management.

Health portraits support various NCD management functions, including risk management (“risk” and “prevention”), diagnosis and treatment (“diagnosis,” “incidence,” and “prediction”), disease progression and mortality stratification (“classification,” “prediction,” and “mortality”), health monitoring (“monitoring,” “tracking,” and “screening”), and providing tailored recommendations (“recommender,” “knowledge,” and “interest”).

High-frequency terms such as “individualized,” “personalized,” “efficacy,” and “cost” emphasize 2 primary goals: addressing unique patient needs and ensuring practical, cost-effective care. Additionally, terms like “information,” “device,” “EHRs,” and “dataset” indicate that health portrait modeling integrates data from diverse sources, including Internet of Things devices, wearables, digital platforms, EHRs, public datasets, and large language models (LLMs). This rich data foundation enables the extraction of relevant features and the creation of digital labels for comprehensive, individualized health portraits.

Through word frequency analysis and content summarization, this study identified 4 primary types of health portraits—diagnostic, prognostic, monitoring, and recommender—which collectively cover the health management continuum for NCDs. The different types of health portraits and their descriptions are shown in Table 2.

Data usage is shown in Figure 3. Overall, only 17.78% of the included studies fulfill the requirements of big data based on the 3V framework. In terms of volume, structured data is highly used across the 4 types, with the lowest rate still reaching 64.29% in recommender portraits. However, the usage of unstructured data (43.33%) shows substantial variation: monitoring portraits lead with 93.33%, while prognostic portraits are at the low end with just 19.05%. Diagnostic portraits, focused similarly on clinical prediction models, exhibit a moderate usage of unstructured data at 42.11%. Regarding velocity, the reliance on digital interactive data (33.33%) is polarized. Diagnostic and prognostic portraits show lower usage rates, at 15.79% and 7.14%, respectively, whereas monitoring and recommender portraits both exceed 85%. For variety, only 31.11% of health portraits encompass all 3 attributes. However, recommender portraits demonstrate the highest performance, with 64.29% of studies encompassing all 3 attributes, among which domain attribute coverage is at 100%. Further, the domain attribute of recommender portraits primarily comes from preference-demand information, which constitutes 78.57% of its data—significantly higher than that in other portrait types. In contrast, prognostic portraits show the most limitations in the domain attribute. Although they excel in specialized medical data, reaching 100% in specific attribute coverage, they exhibit substantial gaps in preference-demand and contextual information, with usage rates of only 2.38% and 4.76%, respectively.

A comprehensive capability assessment was conducted in Figure 4. Overall, the external validation of the big data–driven health portraits is suboptimal, with only 27 studies (30.34%) meeting the required standards. Assessed together with each study's performance in meeting the 3V criteria of big data, only 9 (10.11%) studies meet these standards. Among the 4 types, recommender portraits showed relatively better results, with 4 (28.57%) studies meeting the comprehensive criteria. Finally, we conducted a comprehensive scoping review that synthesizes evidence from diverse literature on big data–driven health portraits, defining the comprehensive scope in Multimedia Appendix 7.

Our findings indicate that big data–driven health portraits primarily cover the full spectrum of NCD management, categorized into 4 types, including diagnostic, prognostic, monitoring, and recommender portraits, each tailored to specific management scenarios. Despite the diversity of data sources for health portraits, a key limitation is the underuse of unstructured and interactive data, which restricts the potential of big data–driven approaches. A quantitative evaluation based on the “3V” framework further demonstrates weak comprehensive capabilities of big data–driven health portraits, largely due to insufficient use of diverse data types. Moreover, the lack of external validation metrics hampers the assessment of the generalizability of health portraits, and there is considerable room for improvement in validation efforts. This review emphasizes the need for continued exploration to enhance the adaptability and effectiveness of big data–driven health portraits in NCD management. The following sections address the key challenges, opportunities, and future roadmap, paving the way for innovative approaches tailored to specific management contexts.

The results showed that 4 primary types of health portraits span the care continuum from hospital-based to home-based settings. Specifically, diagnostic and prognostic portraits are predominantly deployed in clinical settings, where they leverage structured medical metrics and experimental data aligned with traditional health care models [62]. In contrast, monitoring and recommender portraits are increasingly applied in community and home-based contexts, supporting patient self-management, familial support, and interactions with health coaches [63,64]. Each type faces distinct data challenges. Hospital-based portraits benefit from high-quality data but lack continuity, while home-based systems provide continuous input via wearables and mobile devices, often with variable data quality [65]. Monitoring and recommender portraits, in particular, require diverse, high-quality training datasets and rigorous validation to ensure label accuracy and system reliability. Importantly, these types are not mutually exclusive—monitoring and recommender portraits offer real-time responsiveness, while diagnostic and prognostic portraits enhance predictive accuracy. Their integration presents opportunities for more comprehensive, adaptive health management. Emerging technologies, especially the integration of LLMs and knowledge graphs (KGs), are advancing this vision [14,66-70]. LLMs possess robust capabilities for processing unstructured and interactive data, such as chatbot dialogues [71], to extract real-time behavioral embeddings, while KGs structure multimodal biomarkers and social determinants into semantic networks and enable knowledge reasoning based on multimodal information. It must be recommended that this synergy will improve temporal resolution and contextual precision. Such architectures transcend traditional single-dimension health portraits, enabling adaptive health management across biological, behavioral, and environmental tiers [66,67,71]. Numerous experimental studies have shown that integrating LLMs with user historical interaction and behavioral information extracted from KGs can effectively perform relevant predictions and information recommendations. This method has exhibited superior effectiveness and reliability compared to traditional machine learning models across multiple public datasets [72-74]. However, significant challenges persist. Privacy and security concerns are particularly pressing, especially regarding the potential for data breaches during personal data interactions [71]. Addressing these concerns necessitates multifaceted approaches, including blockchain technology, end-to-end encryption, and advanced privacy protocols [75].

This study also reveals that while big data–driven health portraits demonstrate moderate performance across individual dimensions of the 3V framework, their overall effectiveness diminishes sharply when evaluated across all 3 dimensions simultaneously (16 (17.98%) studies). This suggests that health portraits are still in the early stages of development, where the comprehensiveness of data usage remains underdeveloped, which may constrain big data–driven health portraits in patent design or practical application. Upon closer inspection, there are considerable disparities in data use across the 4 types of health portraits. Diagnostic and prognostic portraits show significant differences in the use of unstructured data (volume). For instance, diagnostic portraits depend heavily on unstructured data like high-dimensional imaging, which encapsulates valuable predictive information [76-78], aligning with the findings of Esteva et al [79]. However, the prognostic model mostly used data from public databases, which provided structured self-reported data, though research has reported that integrating unstructured clinical text with structured data can improve model performance [80]. The possible reason for the problem is that, unlike the relatively clear diagnostic conditions and the relatively low requirements for data continuity, the prediction of prognostic outcomes depends on a large number of cohort data before and after treatment, and it is relatively difficult to obtain data from a single study in the real world, especially under the limited patient adherence [81-84]. Our findings also show that monitoring and recommender portraits excel in data volume, transmission, and variety, which may underscore the potential of wearable and contactless devices in health data capture [85-88] as a cost-effective way for all types of portraits to bridge the gaps in acquisition capability of unstructured and interactive data [80]. Moreover, from a portrait integration perspective, as previously discussed, the integration of KGs and LLMs offers a robust framework for achieving multimodal data fusion and health portrait unification [66,67,71]. The maturation of hardware devices supports cross-modal data interoperability, while the development of time-series fusion algorithms transforms conventional 2D labeling systems into dynamic 3D architectures [89,90]. By incorporating temporal granularity, our approach refines population-level subgroup features into individualized trajectory-aware markers, capturing health progression patterns from clinical to community settings. This evolution facilitates the organic unification of the 4 portrait types, thereby enabling precision health management plans tailored to individual developmental trajectories.

Our findings indicate that only 10.11% (9 studies) of studies meeting the “3V” criteria have undergone the external validation. This highlights a significant gap between theoretical potential and practical implementation, underscoring the need for more external validations in the real world. Unlike Youssef et al [91], who cautioned against the reliability of single models validated on limited datasets, our study emphasizes the importance of integrating external validation within a comprehensive big data framework. External validation is essential for assessing model generalizability and performance, supported by prior research [92-94]. We also found that diagnostic and prognostic models emphasize externally validated results [94]. One possible reason for this is the strong association between external validation and study quality evaluation recommended by The BMJ’s guidelines according to the “Prediction Model Risk of Bias Assessment Tool” (PROBAST) [95]. Another possible reason is that the reliability of the prediction results in the real world is the core of the value of model research, aiming to promote results-oriented resource allocation according to accurate risk levels to reduce costs and increase efficiency. Otherwise, models that lack external validation are at high risk of failure in real-world practice. Even though genotype metrics are considered a key component of precision medicine, with advantages such as high standardization, stability, and potential for causal inference, large-scale retrospective studies of predictive models based on genotype metrics have often failed [96]. This may be attributed to population heterogeneity and gene-environment interactions, which can diminish model performance in diverse populations in the real world [97]. Additionally, evaluations of 31 predictive models related to COVID-19 revealed that most studies had a high risk of bias [98]. Consequently, few are used or disseminated in clinical practice, thus failing to promote practical applications [99]. Monitoring and recommender portraits performed relatively poorly in external validation, probably because the guidance of these 2 types of models for health management was more process-oriented, such as patient intervention adherence or facilitating patients' resource usage in limited resources, so they focused more on data transmission speed and update frequency for the accuracy of external validation results in real-time monitoring [65,100]. Recent research has highlighted that intelligent predictive models extending beyond radiology and bioinformatics possess unique characteristics, which may need new evaluation tools such as PROBAST-AI, TRIPOD-AI (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis–Artificial Intelligence) guidelines, and the criteria for health care conversation powered by AI [101,102]. Thus, future studies are recommended to use appropriate standards to complete external validation in real-world data to clarify the reliability of monitoring results and recommendations.

To the best of our knowledge, this study is the first to conduct a scoping review of health portraits from a big data perspective, addressing a critical research gap and laying a solid theoretical foundation for future investigations. One of the key contributions is the introduction of the 3V framework, which operationalizes big data principles to quantify data usage, offering a novel perspective for evaluating health portrait development. Another contribution is that the structured analysis delineates the current state of the field and provides actionable strategies for theoretical advancements and practical applications. Furthermore, compared to traditional health portraits, this scope of big data–driven health portraits provides us with an opportunity to leverage multimodal data while grasping personal dynamic health status adaptively. Despite these contributions, several limitations must be acknowledged. First, restricting the review to English and Chinese studies may introduce language bias, potentially excluding valuable insights from other languages. Second, using a single binary classification metric for external validation limits the evaluation scope, making it difficult to pinpoint deficiencies in studies that do not meet benchmarks. Third, due to limitations in time and resources, this study lacks a more detailed analysis of health portrait features incorporating disease classifications, which will aid in developing a more comprehensive health portrait labeling system.

This study underscores the challenges of big data–driven health portraits in NCD management in Figure 5, highlighting the possible future solutions that can fully realize the transformative potential. Among them, 2 main steps, according to our results, need to be highlighted: (1) prioritize LLM-driven integration of functionalities, contextual applications, and multisource data, while embedding ethical, privacy, and cost considerations at the design stage; and (2) promote more robust external validations according to appropriate standards. Notably, hyperpersonalization, which integrates multiomics data for tailored therapies, is highly recommended as a part of integrated health portraits [103-105]. With its convincing power in uncovering disease mechanisms, it brings the potential to integrate precision medicine with lifestyle nursing based on big data. Additionally, from a health equity perspective, future research could apply the PROGRESS-Plus (place of residence, race/ethnicity/culture/language, occupation, gender/sex, religion, education, socioeconomic status, and social capital) framework to health portraits, generating equity-based labels from natural attribute indicators to promote more equitable resource allocation and care for vulnerable populations [106]. Finally, building on our findings and previous research [18], we propose an integrated detailed roadmap for future practice as shown in Multimedia Appendix 8.

Big data–driven health portraits offer significant potential to enhance personalized and precise management of NCDs. With the hardware advancements and cross-domain collaborations for holistic portrait design, chances are that the transformation of conventional 2D labeling systems into dynamic 3D architectures of integrated big data–driven health portraits through AI-driven approaches will reveal new opportunities. However, future research should focus on the privacy-utility tradeoffs and ethical dilemmas. Addressing these gaps will advance the development of personalized health management solutions.