Introduction

Sleep disorders impose substantial population-level burdens through increased cardiometabolic and neuropsychiatric risk, reduced productivity, and higher healthcare utilization. Because symptom expression and pathophysiology are heterogeneous, multimodal sleep analysis, integrating physiological signals, behavioral data, and clinical context, offers a scalable route to earlier detection, individualized therapy, and system-level efficiency. Sleep is essential for maintaining both physiological and psychological health. However, the global prevalence of sleep disorders continues to rise significantly, posing substantial public health challenges1. According to the World Health Organization’s 2021 survey, nearly one-third of the global population suffers from varying degrees of sleep disturbances. Specifically, in China, the 2025 China Sleep Health Survey highlighted that approximately 48.5% of adults report sleep-related issues, underscoring a pressing public health concern. Sleep disturbances significantly diminish individuals’ quality of life and correlate strongly with increased risks of chronic diseases, including depression, anxiety, obesity, cardiovascular diseases, and Alzheimer’s disease2,3. Additionally, the economic implications are profound, with productivity and healthcare-related losses reaching billions of dollars annually in countries such as the United States. The intricate interplay between sleep disorders, health outcomes, and economic impacts underscores an urgent need for innovative, comprehensive healthcare interventions.

The increasing prevalence of sleep disorders poses considerable challenges to China’s existing healthcare infrastructure. Although demand for specialized sleep healthcare services continues to grow, the distribution of experienced sleep medicine specialists remains highly uneven, disproportionately concentrated in urban tertiary hospitals. According to China’s National Health Commission’s 2023 Statistical Bulletin, only 0.25% of medical institutions nationwide qualify as tertiary or specialized sleep facilities, severely restricting access to essential sleep health services, especially in rural and remote regions. Additionally, traditional clinical management approaches often lack patient-centric interventions, resulting in reduced patient adherence and increased risks of medication dependency, underlining the need for innovative, patient-centered solutions4. Thus, there is an acute need for personalized and effective strategies to enhance healthcare accessibility and quality.

Artificial intelligence (AI), particularly expert-emulating intelligent systems, has emerged as a critical strategy for addressing gaps in healthcare provision, improving diagnostic accuracy, and enhancing therapeutic outcomes5,6. Recognizing the urgency of the sleep health crisis, the Chinese government has prioritized developing specialized psychological and sleep healthcare services, creating a robust policy environment conducive to AI-driven healthcare innovations. This strategic support underscores the government’s commitment to improving national health outcomes and addressing the public health implications of inadequate sleep management.

AI-driven expert agents represent groundbreaking innovations capable of simulating the sophisticated clinical reasoning of specialist practitioners and delivering tailored, comprehensive patient care7,8. Importantly, these systems act as clinical decision support tools rather than autonomous diagnostic devices, augmenting clinician capabilities while maintaining human oversight for high-risk decisions. Aligning with governmental initiatives facilitates the adoption of advanced technology-driven solutions within psychological and sleep healthcare practices, significantly transforming patient engagement and clinical outcomes. Integrating AI-driven interventions into sleep healthcare systems thus represents a critical step towards alleviating public health burdens and achieving equitable healthcare delivery throughout China.

Current landscape and advances

AI is fundamentally reshaping sleep medicine, catalyzing a paradigm shift from reactive treatment to proactive, personalized health management9. This transformation is propelled by twin pillars of technological innovation: advanced deep learning models for physiological signal analysis and conversational AI for human-centric interaction10. Together, these technologies have fostered a diverse product ecosystem, from clinical-grade diagnostic tools to consumer wellness applications. However, this landscape is largely fragmented, highlighting a critical need for integrated, all-in-one platforms like “Hang Hao Meng” that can steward patients along the full continuum of care.

Technological foundation: from automated scoring to predictive intelligence

The early AI applications in sleep medicine focused on overcoming the bottleneck of manual polysomnography (PSG) analysis. Deep learning models, particularly those using Convolutional and Recurrent Neural Networks (CNNs, RNNs), successfully automated sleep staging with accuracy comparable to human experts11. For example, SeqSleepNet12 uses hierarchical recurrent neural networks for end-to-end sleep staging, achieving high performance on a large public dataset. More recently, Transformer-based architectures have further enhanced performance by capturing long-range dependencies in full-night physiological data13.

Despite these advances, challenges such as handling heterogeneous physiological signals and accounting for variance among individuals remained. This spurred the development of more sophisticated architectures. For instance, the Squeeze-and-Excitation Network with Domain Adversarial Learning (SEN-DAL) was proposed to adaptively fuse multimodal data while learning subject-invariant features, thereby improving model generalization14. The most significant leap, however, has been the advent of large-scale, self-supervised foundation models. These models represent a paradigm shift from task-specific training to creating versatile, task-agnostic representations of sleep data. A prime example is PFTSleep, a foundational transformer trained on nearly 14,000 studies from multiple cohorts, which demonstrated state-of-the-art sleep staging accuracy and robust generalization across several diverse, independent test sets15. Further extending this paradigm, the pioneering SleepFM model moves beyond diagnostics to true predictive intelligence. Trained on over 580,000 h of multimodal data, SleepFM generates unified embeddings that can forecast the risk for over 130 future diseases, elevating sleep analysis to a powerful tool for systemic health screening16. While these foundation models offer unprecedented accuracy and predictive power, their dependency on lab-grade PSG data currently limits their scalability.

In parallel, Large Language Models (LLMs) have enabled the creation of conversational agents capable of delivering psychotherapeutic interventions, such as Cognitive Behavioral Therapy for Insomnia (CBT-I). The primary challenge, however, is the inherent risk of “hallucinations”—generating clinically inaccurate advice17. To mitigate this, Retrieval-Augmented Generation (RAG) has become essential. By anchoring LLM responses to a verified knowledge base of clinical guidelines and medical literature, RAG significantly enhances the accuracy and safety of AI-driven advice18,19.

Product ecosystem: global benchmarks and regional innovations

Driven by these technological advances, yet constrained by factors like data dependency and clinical validation, a vibrant global product ecosystem has emerged. Sleepio20 stands as a global benchmark for digital therapeutics (DTx). It offers an automated, 6-week CBT-I program whose efficacy is backed by extensive randomized controlled trials (RCTs), leading to recommendations from NICE in the UK and approval from the FDA in the US. Other digital solutions, including the SleepCare application and Insomnia Coach, have similarly demonstrated efficacy superior to baseline measures in post-treatment follow-up evaluations across RCTs, as noted in systematic reviews assessing mobile self-management tools for insomnia21,22. In the management of obstructive sleep apnea (OSA), ResMed23 has successfully transitioned key aspects of care from clinical settings to at-home environments. Its integrated ecosystem incorporates diagnostic devices and artificial intelligence-driven applications designed to enhance patient adherence to treatment protocols. Nonetheless, these medical-grade solutions primarily address singular sleep disorders, underscoring a persistent gap in comprehensive sleep health management.

Parallel to medical-grade interventions, a diverse consumer wellness market has developed, focusing largely on general sleep hygiene and relaxation rather than clinical conditions. Applications like Sleep Cycle utilize patented acoustic algorithms to analyze sleep patterns and offer smart alarm functions24. Physical devices, such as the Somnox Sleep Robot, even employ biomimicry to physically regulate a user’s breathing rhythm25. However, evidence supporting the consumer-oriented products frequently lacks the robustness and scientific rigor characterizing medical-grade interventions, creating an evident divide between validated clinical therapies and popular wellness-oriented products. This fragmentation contributes to user confusion in distinguishing clinically validated solutions from lifestyle enhancement products.

In China, the sleep health sector has experienced rapid innovation, accelerated by national initiatives including the “Healthy China 2030” framework. Notable among recent developments is the Sumian Pulsed Magnetic Therapy System (PMTS), which integrates pulsed magnetic therapy with digital CBT-I. This approach has demonstrated significant improvements in sleep metrics in multicenter RCTs26. Additionally, NeuroPal, a chatbot- based platform combining chronotherapy with CBT techniques, reported preliminary associations with reductions (approximately 37%) in PSQI scores in early-stage trials27. Despite these promising advances, many regional solutions face limitations such as reliance on single-arm observational data, restricted diagnostic coverage, insufficiently rigorous validation, or fragmented system integration.

Summary of AI powered Sleep Health Applications are listed in Table 1. Collectively, these developments reveal a global product ecosystem characterized by significant potential yet marked by disparities in clinical validation and comprehensive integration. Addressing these challenges will be essential for advancing sleep health solutions that are both scientifically rigorous and accessible to broader populations.

Table 1 Summary of representative AI-powered sleep health applications
Full size table

Next frontier: toward integrated platforms in AI-driven sleep health

While impressive, the current product landscape is highly fragmented. A patient’s journey—from screening and diagnosis to consultation and long-term management—is often fractured across multiple, disconnected single-purpose tools. A patient might use a ResMed device for OSA diagnosis while simultaneously using the Sleepio app to treat comorbid insomnia, creating an inefficient and burdensome experience. Therefore, the next frontier in digital sleep health lies in comprehensive integration. Future platforms must be multimodal in two distinct senses: first, they should be capable of synthesizing a wide range of data modalities—such as conversational text, physiological signals, and clinical reports—and second, they should deliver an end-to-end suite of services, encompassing screening, virtual consultation, digital therapeutics, and personalized treatment planning. Despite rapid progress, most solutions target isolated segments of care (e.g., staging or single-modality CBT-I) rather than integrated, end-to-end management. This gap motivates the design of “Hang Hao Meng”, which unifies screening, diagnostic reasoning, and therapy within a single multimodal agent.

Trustworthiness is foundational for clinical AI. International consensus frameworks (e.g., FUTURE-AI) outline requirements for fairness, consent, explainability, accountability, and regulatory compliance28. We adopt these principles and detail system-level safeguards and evaluation plans in the Trust, Safety, and Governance section.

Hang Hao Meng: a multimodal sleep health expert agent

Illustrative use cases

Use case 1: Insomnia tracking and triage. A user reports difficulty initiating sleep. The agent administers Pittsburgh Sleep Quality Index (PSQI) / Patient Health Questionnaire-9 (PHQ-9) / Generalized Anxiety Disorder-7 (GAD-7) and analyzes free-text inputs. With high PSQI but low mood/anxiety scores, it initiates digital CBT-I, monitors weekly sleep diaries, and adapts guidance to adherence and response. Use case 2: Potential sleep-apnoea screening. Wearable oximetry and heart-rate patterns indicate nocturnal desaturations despite ‘daytime fatigue’ as the chief complaint. The agent flags elevated OSA risk and recommends in-person evaluation and PSG rather than initiating CBT-I.

Techniques integrated

Addressing these challenges head-on, “Hang Hao Meng” Sleep Health Expert Agent was developed collaboratively by the sleep research team at the Mental Health Center affiliated with Zhejiang University School of Medicine and Alipay Healthcare. This sophisticated AI-powered system leverages state-of-the-art generative AI methodologies, capturing and replicating the clinical expertise, extensive knowledge, and nuanced diagnostic reasoning characteristic of sleep medicine specialists. Unlike conventional chatbots, this advanced agent integrates diverse technological components, including LLMs5,29, natural language processing (NLP), privacy-preserving computing frameworks30, comprehensive knowledge graphs31, and multimodal data analytics. Through this integration, the agent delivers a full spectrum of personalized sleep health services, encompassing sleep profiling, intelligent screening, virtual consultation, digital therapeutics, auxiliary diagnosis, and individualized treatment recommendations. A consolidated glossary of abbreviations and technical terms is provided in Supplementary Table S1.

The technical architecture of the “Hang Hao Meng” Sleep Health Expert Agent (Fig. 1) integrates a data foundation, an AI infrastructure, and a model-capability layer that together enable the core functionalities. The Model Capability Layer comprises four key abilities: Question-Answering, Multimodal Analysis, Deep Thinking, and Service Planning. The Question-Answering capability includes intent understanding analysis, medical queries, intelligent health education, and smart triage or pre-diagnosis. Multimodal Analysis handles clinical report interpretation, EEG signal processing, medical imaging analysis, and clinical audio recognition. Deep Thinking supports multi-round complex diagnostics, disease inference, treatment recommendations, and personalized sleep training plans. Service Planning manages appointments, payments, inquiries, interface interactions, dynamic scheduling, and multi-step reasoning tasks. BayLing-Med32 underpins Massive Medical Knowledge for the Question-Answering Ability. Qwen-VL33 supports Full Modality Understanding, crucial for the Multimodal Ability. DeepSeek-R134 enables Deep Reasoning, powering the Deep Thinking Ability, and supports Complex Planning for the Service Planning Ability. These models allow the agent to handle complex interactions and tasks.

Fig. 1: Architecture of “Hang Hao Meng” Sleep Health Expert Agent.
Fig. 1: Architecture of “Hang Hao Meng” Sleep Health Expert Agent.
Full size image

The architecture consists of four layers: the Data Foundation Layer, AI Infrastructure Layer, Model Capability Layer, and the Agent Application Layer. The Data Foundation Layer integrates heterogeneous sources, including clinical data, user behavior from Alipay, wearable device data, and medical guidelines or expert consensus. The AI Infrastructure Layer supports scalable computation and stable deployment with components such as the inference engine (MAYA, AIS), runtime environment (ARECLL), and distributed computing engines (ODPS, Ray, Spark), along with feature management and fallback strategies. The Model Capability Layer comprises three core functional blocks powered by foundation models: BayLing-Med for intent understanding, medical question-answering, triage, and health education; Qwen-VL for EEG signal interpretation, multimodal analysis, and clinic audio understanding; and DeepSeek-R1 for complex diagnostic reasoning, treatment planning, service orchestration, and multi-interface execution. At the top, the Agent Application Layer delivers end-user services including sleep profile generation, intelligent consultation and screening, digital therapeutics, and personalized diagnosis and treatment recommendations.

The Artificial Intelligence Infrastructure Layer (AI INFRA) establishes the foundational computational and operational framework for agent-based systems, comprising three core components: Inference Engine, ARECLL, and Computing Engine. The Inference Engine subsystem integrates specialized platforms such as MAYA (a proprietary machine learning platform) and AIS (AI Inference Service) for executing machine learning workloads. ARECLL provides a resilient runtime environment featuring Runtime, Flexible Fallback, and Feature Management. Supporting processing requirements, the Computing Engine employs established frameworks including ODPS (Open Data Processing Service, a large-scale computing platform), Ray, and Spark for data processing and computing.

Concurrently, the Data Foundation Layer implements a comprehensive data integration strategy through heterogeneous data source consolidation essential for operational efficacy. This includes Clinical Data, Personal Wearable Devices Data, Alipay User Behavioral Data, and Medical Guidelines & Expert Consensus.

Interoperability and canonical data modeling. Heterogeneous inputs from EHRs, wearables, and patient-reported outcomes are mapped to a canonical data model aligned with HL7 FHIR resources and, where applicable, the OMOP (Observational Medical Outcomes Partnership) common data model. In practice, real-world deployment across different health systems may still be constrained by semantic mapping inconsistencies and licensing requirements. Semantic normalization binds local terminologies to Systematized Nomenclature of Medicine (SNOMED CT), International Classification of Diseases, Tenth Revision (ICD‑10) and Eleventh Revision (ICD‑11), RxNorm (a normalized naming system for clinical drugs developed by the U.S. National Library of Medicine), and Logical Observation Identifiers Names and Codes (LOINC) via a managed terminology service. Institution-specific ETL profiles handle source ingestion, validation, and conformance checks; the AI layer consumes only the canonical schema, enabling reuse of clinical logic and retrieval prompts across sites. For new deployments, only the ETL profile and terminology bindings require adaptation, while inference, triage logic, and RAG retrieval remain unchanged. This separation reduces integration friction, improves auditability, and supports portability across healthcare domains.

At the core is a domain-specialized conversational AI optimized for multi-turn diagnostic dialogue. We initialize with a high-capacity foundation model and align it to clinical practice in two phases. First, supervised fine-tuning with chain-of-thought exemplars instills stepwise clinical reasoning, including hypothesis generation, evidence weighing, and next-question planning. Second, a reinforcement phase termed Reinforcement Learning with Verifiable Labels uses programmatically generated labels from our medical knowledge graph to encode diagnostic categories and severity against authoritative standards such as ICD-10. This verifiable reward signal reduces susceptibility to subjective feedback and supports safe optimization of diagnostic behavior.

At inference time (Fig. 2), each conversational turn is grounded by a RAG pipeline that fetches guideline-level evidence and salient literature from the knowledge graph. The retrieved context conditions the LLM’s chain-of-thought reasoning, which iteratively narrows differential diagnoses and selects the next best question until predefined diagnostic thresholds are met. The same mechanism then retrieves guideline-concordant care recommendations. This design provides a transparent audit trail. By coupling guideline citations with a clinician-facing reasoning summary (implemented and available during the review period), the system supports auditability and helps mitigate ungrounded outputs, facilitating clinician oversight.

Fig. 2: Dual-phase model training and iterative inference workflow of “Hang Hao Meng”.
Fig. 2: Dual-phase model training and iterative inference workflow of “Hang Hao Meng”.
Full size image

Schematic of a specialized conversational AI for clinical diagnosis using retrieval-grounded, chain-of-thought reasoning. A Training. A large language model (LLM) is adapted with supervised fine-tuning (SFT) using chain-of-thought (CoT) exemplars to instill clinical reasoning, followed by reinforcement learning with verifiable labels (RLVR). RLVR uses a medical knowledge graph to programmatically generate objective diagnostic labels (for example, ICD-10 criteria) as reward signals, promoting evidence alignment and reducing reward hacking. B Inference. During multi-turn dialogue, retrieval-augmented generation (RAG) fetches guideline-level evidence from integrated knowledge sources; the LLM performs CoT to analyze symptoms, refine differentials, and ask targeted follow-ups until diagnostic thresholds are met. Outputs include triage recommendations and guideline-based treatments, with auditable reasoning summaries and source citations.

Further augmenting the agent’s practical capabilities is the integration of digital human or digital twin technology. This innovative component specifically replicates the expertise and clinical decision-making processes of renowned Chinese sleep expert Dr. Mao Hongjing, ensuring the widespread and enduring dissemination of expert-level knowledge and practice. To preserve evaluation independence, diagnostic concordance was reviewed by clinicians who were not involved in system development. By systematically integrating multimodal datasets—including comprehensive medical histories, detailed imaging results, recorded audio consultations, structured electronic clinical records, wearable health-monitoring data, and behavioral insights derived from platforms such as Alipay—the agent achieves highly precise, individualized patient assessments. This holistic approach accurately mirrors the nuanced evaluations typically performed by expert clinicians, significantly enhancing patient outcomes.

To ensure the provision of scientifically valid and evidence-based information, particularly within the Q&A capabilities, the system leverages techniques like RAG35,36. These advanced techniques mitigate common AI limitations, such as inaccuracies and so-called “hallucinations”37, by meticulously curating extensive clinical information, scientific literature, and authoritative medical guidelines into structured, comprehensive knowledge graphs, consistently delivering credible clinical recommendations. Complementing these clinical decision-support tools, the agent integrates digital therapeutic interventions tailored explicitly for sleep disorders. Leveraging advanced software applications, wearable devices, mobile health technologies, and immersive virtual reality solutions, the system delivers targeted CBT-I, personalized relaxation techniques, structured sleep education programs, and continuous patient monitoring, thus comprehensively addressing various dimensions of sleep health management.

Ethical considerations and real-world deployment

This work reports an observational program evaluation of the “Hang Hao Meng” sleep health expert agent. The study was approved by the Ethics Committee of Hangzhou Seventh People’s Hospital (Approval No. 2024-044). Before accessing the service, users must complete a mandatory click-wrap agreement that describes the nature of the service, data categories collected, intended uses, and privacy protections. To ensure transparency, that users are interacting with an AI system rather than receiving clinician-delivered care, the interface displays a persistent disclosure stating that the content is AI-generated, does not constitute a medical diagnosis, and is not a substitute for professional medical advice; users are advised to seek in-person care if symptoms worsen or urgent concerns arise. The agent also enforces fail-safe safety rules: when inputs indicate severe distress or potential self-harm risk, it withholds treatment recommendations and prompts immediate escalation to offline clinical services.

For the evaluation, only aggregated and de-identified records from collaborating clinical sites were included. The analytic cohort represents a clearly defined subset of all users who accessed the service. These data were used exclusively for observational assessment and quality improvement purposes, with no experimental manipulation of care pathways.

The deployment of the “Hang Hao Meng” Sleep Health Expert Agent has demonstrated significant feasibility and operational utility across multiple healthcare contexts. Through advanced online consultation and intelligent triage capabilities, the system offers continuous, 24 h virtual consultations, initial risk assessments, and targeted referrals. Clinical implementation data as of June 30, 2025, indicate substantial impact: the system has served over three million users, identifying 90,100 individuals with sleep disorders. For this population, online and offline medical services were provided. Mild cases received 70,200 online interventions, while 20,200 severe patients were referred for offline consultations. This stratified approach optimizes resource allocation and ensures timely intervention.

To quantify real-world performance, we analyzed operational data from October 2024 to August 2025 with prespecified endpoints covering diagnostic concordance, patient-reported outcomes, and engagement. We note that these findings are derived from real-world observational data and should be interpreted as hypothesis-generating; they do not establish causal effects. The performance data presented here are derived from internal operational audits conducted within collaborating institutions. To enhance generalizability and reduce assessment bias, we plan external validation and clinician-blinded studies at independent sites.

Diagnostic concordance

Among 441 cases with clinician follow-up, we evaluated agreement between the agent’s top-3 ICD-10 major-category suggestions and physicians’ primary diagnoses. Concordance—defined as ≥1 overlapping label between AI suggestions and clinician diagnoses—was 78.7% (347/441). We report top-3 concordance to reflect the agent’s role as a clinical decision-support aid that surfaces a differential shortlist, rather than as a standalone diagnostic system. Performance varied by condition prevalence: accuracy reached 100% for sleep disorders (ICD-10 F51, n = 143), 92.0% for anxiety disorders (F41, n = 162), 72.9% for depressive episodes (F32, n = 48), and 22.7% for other combined categories (n = 88). The 441 cases were randomly sampled from the clinical records of the Sleep Medicine Center of Hangzhou Seventh People’s Hospital to reduce potential selection bias. The sample was not enriched for diagnostic complexity and reflects the case mix of this specialist clinic during the study period. Figure 3A shows diagnostic distributions across four aggregated categories (F51 comprising 32.4% and F41 36.7% of cases). Figure 3B details per-category concordance, and Fig. 3C (Sankey) visualizes the principal error modes, including frequent confusion between F51 and F41 despite high within-category concordance for both. These patterns highlight challenges in differentiating symptomatically overlapping mental health and sleep conditions.

Fig. 3: Diagnostic distribution and agent–clinician concordance.
Fig. 3: Diagnostic distribution and agent–clinician concordance.
Full size image

A Distribution of physician-assigned diagnoses (n = 441) across aggregated ICD-10 categories: F51 (sleep disorders; n = 143), F32 (depressive episodes; n = 48), F41 (anxiety disorders; n = 162), and OTHER (combined; n = 88). B Per-category concordance for the agent, defined as the proportion of cases in which at least one of the agent’s top-3 ICD-10 categories matched the physician’s primary diagnosis: F51 100.0%, F32 72.9%, F41 92.0%, OTHER 22.7%. C Sankey diagram mapping physician labels (left) to the agent’s top-3 categories (right); flow width is proportional to case count. Green flows indicate matches and gray flows indicate mismatches, highlighting frequent cross-classification between sleep and anxiety disorders despite high within-category concordance.

Triage and referral

Triage combines rule-based thresholds with AI guidance using self-reported symptoms, intent recognition for safety phrases, and standardized scales. Typical routing applies PSQI 8–14 to online CBT-I and PSQI > 15 or elevated GAD-7 / PHQ-9 to offline consultation. In particular, for presentations suggestive of obstructive sleep apnoea (e.g., severe nocturnal desaturation patterns) or moderate-to-severe mood or anxiety symptoms with self-harm risk, the agent is explicitly configured not to provide standalone diagnostic conclusions or medication recommendations; instead, it prioritizes risk flagging and expedited referral to in-person sleep or psychiatric evaluation. Full “triage success” confirmation requires longitudinal outcome linkage; this analysis reports preliminary operational performance and flags prospective validation (see Limitations).

Patient outcomes and engagement

Over the follow-up period, participants showed clinically meaningful improvement across all domains (Fig. 4; lower scores indicate symptom improvement). Confidence intervals for mean changes used user-level t-based estimates, effect sizes used pooled baseline–endline standard deviations (SD), and 95% confidence intervals for effect sizes were obtained via 1000-resample percentile bootstrapping. Sleep Quality (PSQI) decreased from 12.12 (SD 5.01) at baseline to 9.20 (SD 4.46) at endline, corresponding to a mean reduction of −2.91 points (95% CI − 3.18 to −2.64) and a standardized effect size of Cohen’s d = 0.61 (95% bootstrap CI 0.56–0.68). Depressive symptoms (PHQ-9) declined from 7.16 (SD 6.72) to 4.04 (SD 5.21), a mean reduction of −3.12 points (95% CI − 3.45 to −2.80) and a medium effect size (Cohen’s d = 0.52; 95% bootstrap CI 0.47–0.58). Anxiety (GAD-7) decreased from 6.05 (SD 6.21) to 2.98 (SD 4.46), a mean reduction of −3.07 points (95% CI − 3.38 to −2.76) with Cohen’s d = 0.57 (95% bootstrap CI 0.51–0.63). Because endline was defined as the last observed assessment, these estimates may be sensitive to missingness patterns; future prospective evaluations will prespecify longitudinal models (e.g., mixed-effects models) with sensitivity analyses. During 8 Aug–4 Sep 2025, engagement averaged daily active user 23,480, page views 37,211, and interaction depth 1.66. Cohort-based retention among new users initiating the agent was 2.28% (day 1), 11.85% (day 7), and 2.67% (day 30). Engagement metrics were consistent across agent variants, with variation <2.00% for daily active users and <3.00% for retention.

Fig. 4: Longitudinal trends in patient-reported outcomes.
Fig. 4: Longitudinal trends in patient-reported outcomes.
Full size image

Spline-smoothed mean trajectories for PSQI, GAD-7, and PHQ-9 over 335 days in users enrolled in the CBT-I program, with shaded bands denoting the standard error of the mean.

Clinical workflow impact

Pre-visit data synthesis generates structured reports that accelerate clinician review by consolidating histories, wearable metrics, and sleep diaries into a single summary. Preliminary time-motion observations at Hangzhou Seventh People’s Hospital suggest an initial consultation time reduction of ~20–30%. These reports also anchor guideline-concordant recommendations, including lifestyle modification, CBT-I protocols, and medication guidance tailored to individual profiles.

Longitudinal care

The platform supports digital rehabilitation and continuous monitoring, enabling dynamic adjustment of therapy according to reported outcomes and sensor trends. This feedback loop sustains engagement and facilitates earlier course-corrections in long-term management.

Perspective-level assessment

As a fully integrated, multimodal sleep agent, “Hang Hao Meng” offers a promising model for future AI-augmented healthcare. Its successful pilot and deployment illustrate how advanced AI can be embedded into the full continuum of care—from screening and risk stratification to consultation and ongoing therapy. The agent’s ability to interact naturally with patients, preserve privacy, and ground decisions in clinical evidence exemplifies how digital health tools can replicate specialist-level care at scale.

Key areas for further development include independent, peer-reviewed clinical trials to assess comparative effectiveness against human-delivered care; evaluation of long-term outcomes across varied clinical cohorts; transparency audits to ensure fairness and bias mitigation; and regulatory frameworks supporting clinical integration and data governance. Furthermore, studies of cost-benefit, clinician acceptance, and user satisfaction will be crucial in ensuring successful national rollout and healthcare system integration.

Although the current implementation is tightly coupled to the Chinese ecosystem, internationalization is feasible with a structured plan. First, interoperability and privacy: align data exchange with HL7 FHIR, adopt SNOMED CT/ICD/LOINC for terminology, and comply with regional privacy regimes such as GDPR, HIPAA, and PIPL using data minimization, audit logging, and, where required, federated learning or differential privacy38,39. Second, model generalization: extend training and calibration with multi-country datasets, local guidelines, and formulary differences40. Third, product localization: adapt content, reading level, and interaction design to cultural norms and languages, and audit subgroup performance41. Fourth, regulatory strategy: pursue SaMD pathways (e.g., FDA 510(k)/De Novo, EU MDR CE, UKCA) with post-market performance monitoring and incident reporting42,43.

Limitations

This Perspective reports early operational experience and interim analyses. The present findings are derived from real-world observational program data and therefore cannot establish causality. Confounding factors, selection bias, and regression to the mean may contribute to observed symptom improvements. The absence of RCTs means current evidence supports operational feasibility and preliminary real-world utility rather than proven clinical efficacy. Formal cost-effectiveness evaluations and adjudicated diagnostic accuracy studies limit generalizability. Triage “success” has not yet been confirmed with long-horizon outcome linkage at scale. In this dataset, a dedicated clinician–clinician inter-rater reliability assessment was not prespecified, so a formal human-reference baseline for interpreting AI–clinician concordance is not yet available. In future work, we will incorporate clinician agreement analyses alongside prospective randomized controlled trials to strengthen the evidentiary basis for both accuracy and clinical impact.

The current dataset originates primarily from Zhejiang Province and may introduce demographic and geographic bias. Among 2,043 users, 82.3% reported lower educational attainment; adults ≥65 years comprised ~8%; and 75.3% were female, consistent with sex differences in sleep-related help-seeking. Medication exposure was common (current 75.2%; past 63.8%), suggesting a clinically severe cohort. To improve generalizability, we will expand to multi-site cohorts across diverse regions, prioritize under-represented subgroups (e.g., elderly, male, medication-naïve), implement fairness audits using demographic parity difference (current estimate 0.51 by sex, indicating disparities in training data or usage patterns that warrant active mitigation), equalized-odds, and calibration across subgroups, and report subgroup-stratified effects by age, education, and geography. Because endline was defined as last assessment, estimates may be sensitive to missingness patterns; future prospective evaluations will prespecify longitudinal models with sensitivity analyses. While these characteristics may limit broad generalizability, they provide insight into digital CBT-I among underserved populations with constrained access to traditional services.

These gaps are acknowledged, and prospective work is planned, including multi-site diagnostic-accuracy studies with adjudicated reference standards, CONSORT-AI–aligned trials44 for comparative effectiveness, economic evaluations, and structured assessments of clinician workload, satisfaction, and equity across subgroups.

Trust, safety, and governance

Principles and scope

“Hang Hao Meng” is being developed in line with international guidance for trustworthy clinical AI, with lifecycle governance across data, models, and deployment28,38,40. The project is a non-profit public-health initiative delivered via mobile platforms. The app is free to download and install, with intelligent consultation and screening provided free of charge.

Data governance and consent

Users complete a digital, plain-language consent that details data categories, purposes, sharing, and rights (access, export, withdrawal); consent status is logged and versioned. Health data are processed only for care delivery, safety monitoring, and model quality improvement under a documented purpose-limitation policy. No health data collected by the agent are used for commercial financial products or services, and the health-data pipeline is strictly firewalled from financial services. Security controls include role-based access control (RBAC), least-privilege permissions, encryption in transit and at rest, and immutable access logging. Periodic privacy audits are conducted to verify segregation and ensure that no data flows into financial systems.

Data retention

Personally identifiable data are pseudonymized; clinical interaction logs and learned features are stored on encrypted, access-controlled servers. Routine retention is 24 months for operational analytics and safety monitoring, after which records are deleted or irreversibly de-identified unless longer retention is required by law or explicit user permission. Users may request early deletion; completion is confirmed and logged. Consent and data governance follow a dynamic-consent model with clear, in-product disclosures of data use, retention, and opt-out options.

Privacy-preserving operation and portability

Privacy protections include data minimization and, where data residency is required, federated learning and differential privacy41. Regulatory compliance and portability are pursued through interoperability with HL7 FHIR and standard terminologies (SNOMED CT/ICD/LOINC), jurisdiction-specific privacy and AI regulations (e.g., GDPR/HIPAA/PIPL), and software-as-a-medical-device pathways (FDA/CE/UKCA), coupled with post-market surveillance.

Model lifecycle and monitoring

Model and knowledge-base updates occur on a scheduled quarterly cycle or ad hoc for critical guideline changes. Each update includes provenance, validation reports, and rollback capability; post-deployment monitoring tracks performance, calibration, and subgroup fairness, with incident management and human-in-the-loop escalation for safety events. Clinical guidelines are encoded in modular knowledge bases, allowing targeted updates without retraining the core models. Updates are versioned with provenance, effective dates, and rollback options, and are reviewed periodically by a clinical oversight committee.

Role separation and oversight

To reduce the risk of sponsor bias and circular validation, we maintain a clear separation of duties across three functions. First, the development function comprises Ant Group engineers and clinical informatics specialists responsible for system architecture, model training, digital-human (“digital twin”) implementation, and platform integration. Second, clinical curation and oversight are provided by a multidisciplinary oversight committee, including senior clinicians such as Dr. Mao Hongjing, who review guideline alignment, approve major protocol changes, and oversee safety policies but do not adjudicate individual validation cases. Third, the evaluation function is performed by independent clinicians at collaborating hospitals who are institutionally separate from the development team and are responsible for real-world deployment monitoring, diagnostic concordance assessment, and outcome evaluation.

Fairness and representativeness

Fairness is monitored via pre-specified subgroup audits, calibration checks, and error-case review, with corrective actions at training and post-processing stages and disaggregated reporting as sample sizes permit39. These controls are maintained throughout the lifecycle and aligned with the governance principles cited above.

Explainability and transparency

Explainability and transparency are supported by retrieval-augmented responses with guideline citations and a clinician-facing reasoning summary that provides an auditable trail; guardrails restrict ungrounded outputs in high-risk contexts42,43. Illustrative reasoning summary: Recommendation: psychiatric evaluation. Rationale: (i) PHQ-9 = 15 and GAD-7 = 12 indicate moderate-to-severe mood/anxiety symptoms; (ii) free-text notes include “persistent low mood” and “anhedonia “, consistent with MDD criteria; (iii) minimal PSQI improvement after 4 weeks of CBT-I; (iv) guideline concordance for insomnia with comorbid moderate-to-severe mood disorder. The summary links recommendations to patient-specific data elements and cites guideline passages retrieved by RAG. Auditable traces and citation lists are currently used for internal safety audits and retrospective review. A real-time clinician interface is scheduled for broad deployment in January 2026.

Fail-safe operation and human oversight

For rare presentations, conflicting signals, or edge cases insufficiently covered by current protocols, the agent enters a fail-safe mode that withholds definitive recommendations, surfaces uncertainty to the user, and triggers immediate referral to in-person care. Responsibility for high-risk decisions remains with licensed clinicians, positioning the agent strictly as a clinical decision support tool. This human-in-the-loop design prioritizes safety and creates a feedback pathway to identify knowledge gaps, informing subsequent curation of the knowledge base and model fine-tuning.

Ongoing commitments

Consistent with the FUTURE-AI principles, we will continue bias analysis and auditing throughout the lifecycle; enhance dynamic-consent UX and disclosures; expand citation precision and empirical tests of interpretability; strengthen quality control for inputs and outputs with clinician confirmation for critical recommendations; and maintain continuous compliance programs for privacy, security, and regulatory readiness.

Conclusion and future outlook

The evolution of AI-powered sleep agents marks a profound transformation in healthcare from traditional reactive models toward home-first, multimodal, phenotypic care. Advances in wearable sensors, signal processing, and LLMs now enable continuous, personalized monitoring and insights—allowing sleep disorders to be identified and addressed early, within daily life. The integration of real-world, multimodal data and adaptive AI not only enhances diagnostic precision but also supports the transition to remote patient monitoring and learning health systems—where data from every user interaction informs continuous model refinement and outcome improvement.

However, the promise of AI in sleep healthcare depends heavily on rigorous clinical evidence, regulatory robustness, and explainable human–AI collaboration models. Emerging guidelines such as FUTURE-AI emphasize the need for fairness, traceability, usability, and robustness in medical AI28. Digital health regulators (e.g., NICE, FDA, MHRA) are increasingly adopting frameworks enabling expedited approval for validated AI applications, provided they demonstrate clinical safety and efficacy45,46. Moreover, human-in-the-loop architectures—where clinicians oversee and guide algorithmic outputs—are being recognized as essential for ensuring patient safety and maintaining trust.

Within this emerging paradigm, “Hang Hao Meng” stands as a pioneering exemplar of scalable, integrated sleep-health systems. By delivering end-to-end care—from risk screening to personalized treatment and follow-up—via multimodal AI grounded in expert knowledge, it embodies the shift toward comprehensive, personalized, and accessible care. Its deployment to over 4 million users with more than 90,000 clinical screenings demonstrates feasibility and scale. To fully realize its potential, formal validation through peer-reviewed trials and long-term outcome monitoring must follow.

Looking forward, frameworks for AI-augmented sleep care must embed four core pillars: 1) equitable access through at-home multimodal monitoring; 2) clinical-grade validation and transparent regulatory compliance; 3) hybrid workflows where AI complements—rather than replaces—clinicians; and 4) continuous learning systems that iterate based on real-world outcomes and user feedback. “Hang Hao Meng” already exemplifies this model, positioning it as a blueprint for next-generation digital therapeutics.

Ultimately, the expansion of AI-driven sleep health platforms like “Hang Hao Meng” across national health systems could significantly reduce disease burden, bridge healthcare inequities, and improve population health—solidifying AI’s role not just as a technical innovation, but as a driver of public-health transformation.