跳到正文
原文
Nature 健康护理主题·· 2026-10-09AI 评分43

拓展 AI 驱动康复运动游戏的功能范围:多模态生物反馈与自适应安全监测

Expanding the functional scope of AI-driven rehabilitation exergames: multimodal biofeedback and adaptive safety monitoring

AI 导读

一篇评论提出,在 Tannús 等 AI 驱动低成本卒中康复运动游戏框架中加入实时心率反馈,构建闭环自适应安全监测,按个体阈值调节难度或触发警报。该框架用普通 RGB 摄像头预测 FMA 评分,12 名参与者(24 条肢体)中 R²=0.89。评论指出,卒中幸存者常伴心血管合并症,居家无监督训练需个体化运动剂量。

正文

We read with great interest the study by Tannús and colleagues, “AI-driven low-cost rehabilitation exergame as a lightweight framework for stroke assessment,” recently published in npj Digital Medicine1. The authors demonstrated that upper-limb motor performance, captured through a standard RGB camera, can predict Fugl-Meyer Assessment (FMA) scores with high accuracy using an interpretable linear model. Their choice to prioritise transparency and low-cost scalability represents a meaningful contribution to digital stroke rehabilitation.

Here, we focus on a single, clinically urgent extension of the Tannús framework: integrating real-time cardiovascular safety monitoring into the exergame’s adaptive loop. While their system accurately assesses motor function, unsupervised home use raises the risk of overexertion in stroke survivors with often-compromised cardiac status. We propose a closed-loop adaptive framework where heart rate feedback modulates game difficulty or triggers alerts, ensuring exercise intensity remains within individually safe limits. This commentary’s sole novel contribution is a technically specified, clinically validated pathway to achieve such closed-loop safety, moving beyond general speculation.

Cardiovascular monitoring and individualised exercise boundaries

The Tannús framework focuses on kinematic assessment of motor performance. However, rehabilitation effectiveness and physiological safety represent complementary dimensions of therapeutic dosing. Stroke survivors frequently present with cardiovascular comorbidities, autonomic dysfunction, or medication-modified heart rate responses.

The American Heart Association/American Stroke Association scientific statement on physical activity for stroke survivors emphasises that exercise prescriptions should be individualised based on cardiovascular status and medication effects2. Age-predicted heart rate zones are commonly used in practice, yet β-blocker therapy and post-stroke autonomic impairment can blunt chronotropic responses.

Furthermore, observational studies have demonstrated that upper-limb task practice in chronic stroke often elicits low cardiovascular load, with only a subset of tasks reaching moderate-intensity thresholds3. These findings suggest that exercise intensity is highly task-dependent and may not be adequately inferred from movement amplitude alone.

In this context, integration of real-time heart rate monitoring into camera-based exergames could enhance safety oversight and allow adaptive dosing. For example, game difficulty could be automatically adjusted when heart rate exceeds individualised thresholds, or alerts could be issued when sustained high-frequency movements occur without proportional cardiovascular adaptation.

Among candidate physiological signals, heart rate (HR) offers the highest feasibility for real-time safety monitoring, as it can be reliably acquired via consumer-grade PPG smartwatches or chest straps. Heart rate variability (HRV) may provide complementary information on autonomic fatigue, though its reliability requires further validation in post-stroke populations. Other signals, such as photoplethysmography amplitude and electrodermal activity, currently lack sufficient evidence for clinical deployment in this context and are not incorporated into our proposed framework.

Multisensory feedback and motor learning optimisation

The following discussion on multisensory feedback is provided as an auxiliary observation and does not form part of our core proposal, which remains focused on cardiovascular safety. Gamification enhances engagement, yet the structure of feedback delivery strongly influences motor learning.

Adaptive visual feedback in virtual rehabilitation environments has been shown to guide patients toward more physiological movement patterns4. In addition, systematic reviews suggest that combined visual and auditory biofeedback can improve motor recovery after stroke, although heterogeneity remains high across intervention designs5. A natural extension of the Tannús system would be to use detected kinematic deviations (e.g., reduced hand opening) to trigger real-time corrective cues, which is technically feasible within the existing RGB camera framework.

From static assessment to closed-loop adaptive rehabilitation

While broader adaptive strategies such as patient stratification represent important future research avenues, they are mentioned here only to highlight complementary directions rather than as elements of our proposed framework.

With 12 participants (24 limbs), Tannús et al. reported an R² of 0.89 for FMA prediction. As sample sizes increase, inter-individual heterogeneity—including compensatory strategies, limb dominance differences, and postural variability—may challenge global model generalisability.

Recent work in stroke rehabilitation has emphasised stratified modelling approaches and patient subgroup identification to improve predictive performance and personalisation6. More broadly, predictive tools in stroke recovery increasingly aim to support individualised rehabilitation pathways rather than static categorisation7.

The key conceptual strength of the Tannús platform lies in the fact that assessment occurs during gameplay. Although our primary proposal concerns cardiovascular safety, the Tannús platform’s assessment-during-play structure also invites other forms of closed-loop adaptation, such as using predicted motor scores to dynamically inform task difficulty, workspace demands, or movement speed requirements. This structure provides a feasible basis for these extensions.

We therefore ask whether assessment outputs are currently used to adjust gameplay parameters in real time, and whether future versions could incorporate stratification-informed adaptive algorithms. These remain open research directions rather than immediate extensions. Such evolution—from passive assessment to responsive rehabilitation intervention—may represent an important direction for digital stroke therapy platforms.

The considerations above are offered in appreciation of the authors’ contribution and in the spirit of constructive scientific exchange. We look forward to further developments of this lightweight and scalable framework.

Feasibility, validation roadmap, and practical challenges

Feasibility—Consumer-grade wearables (e.g., Apple Watch, Fitbit, Polar) provide heart rate data via Bluetooth Low Energy (BLE), which can be received by the existing Unity-based exergame engine to adjust game parameters or trigger safety alerts.

Validation roadmap – Clinical translation of the proposed closed-loop adaptive system should proceed in three phases as illustrated in Fig. 1. Phase 1 involves threshold simulation in healthy subjects, validating alert logic and latency. Phase 2 moves to a laboratory setting with stroke survivors, using electrocardiography (ECG) as the reference standard for heart rate measurement. Phase 3 is a home-based pilot deployment with automated safety monitoring, assessing real-world outcomes such as adverse events, adherence, and usability. The final goal is a safe, scalable, AI-driven home rehabilitation system for stroke survivors.

Fig. 1: Three-phase validation roadmap for the proposed closed-loop adaptive exergame system.

A Closed-loop safety monitoring framework. B Three-phase validation strategy: Phase 1: Healthy subjects – threshold simulation, alert logic, latency validation. Phase 2: Stroke survivors in lab – ECG reference, agreement analysis (Bland-Altman). Phase 3: Home pilot – safety monitoring, adverse events, adherence, usability.

Challenges and mitigations – Motion artifacts can be reduced using accelerometer-assisted PPG or a chest strap. For patients on β-blockers or with pacemakers, target heart rate zones are invalid; the Borg rating of perceived exertion or kinematic changes (e.g., sudden slowing) should serve as proxies. Real-time latency must remain below two seconds for safety (achievable with BLE 5.0). Sensor synchronisation can be managed via timestamp alignment using BLE 5.0 or local network time. TinyML models are mentioned as a forward-looking opportunity rather than a validated component; their integration would require dedicated training datasets and prospective validation in stroke populations, which we do not propose in the present commentary.

Opportunities from current wearable technology: Consumer-grade devices (e.g., Apple Watch Series 8 + , Polar H10) have been reported to meet the latency (<2 s) and accuracy (heart rate within 5 bpm of ECG during moderate exercise) requirements in healthy individuals under controlled conditions. However, we acknowledge that PPG accuracy varies by device model, skin tone, placement, and arrhythmia status; these sources of variability would be systematically quantified against ECG in Phase 2 of our proposed validation

Operational specifications for the proposed validation

To transform the three-phase roadmap into a technically actionable framework, the following operational details should be specified:

Individualised heart-rate thresholds

Thresholds would be derived from age-predicted maximum heart rate (220−age), adjusted by resting heart rate and β-blocker dosage when applicable, with a target zone of 40–60% of heart rate reserve (Karvainen method). Final thresholds would be set by the supervising rehabilitation physician prior to each session based on the patient’s current medication and clinical status.

Alert and difficulty-adjustment logic

A visual and auditory alert would be triggered when heart rate exceeds the individualised threshold for >15 seconds. If elevation persists beyond 30 s, the game difficulty would be reduced linearly (e.g., speed reduced by 25%, or target workspace shrunk). A sudden heart rate drop of >20 bpm within 10 s would trigger an immediate pause and prompt the patient to rest.

False-positive and false-negative handling

A double-confirmation mechanism would be employed: an alert is automatically cancelled if heart rate returns below the threshold within 10 s of the initial trigger. If three consecutive alerts occur within a 15-min window, the system enters a mandatory cool-down period (5 min of rest) and logs the event for remote review.

Safety event adjudication

All triggered alerts, system pauses, and patient-reported symptoms would be reviewed within 24 h by a remote rehabilitation team (therapist or physician). Events would be classified as mild (self-resolved), moderate (required manual intervention), or severe (required medical attention).

Primary and secondary endpoints

The primary endpoints would be: (1) agreement between wearable-derived heart rate and ECG (assessed via Bland-Altman analysis), and (2) alert frequency per session. Secondary endpoints would include system usability score (SUS), patient adherence (session completion rate), and incidence of adverse events.

Statistical methods and sample size

For Phase 2 (laboratory validation), a sample size of 20 stroke survivors would provide 80% power at α = 0.05 to detect a clinically meaningful difference in heart rate agreement (based on a moderate effect size from prior wearable validation studies in stroke populations). For Phase 3 (home pilot), 15 participants are proposed as a feasibility benchmark, following established guidance for pilot studies in rehabilitation technology.

Data availability

No datasets were generated or analysed during the current study.

References

  1. Tannús, J., Valentini, C. & Naves, E. AI-driven low-cost rehabilitation exergame as a lightweight framework for stroke assessment. NPJ Digit Med. 9, 196 (2026).

    Article  PubMed  PubMed Central  Google Scholar 

  2. Billinger, S. A. et al. Physical activity and exercise recommendations for stroke survivors: a statement for healthcare professionals from the American Heart Association/American Stroke Association. Stroke 45, 2532–2553 (2014).

    Article  PubMed  Google Scholar 

  3. Davidson, S. et al. Characterizing heart rate response during upper extremity repetitive task practice in chronic stroke. Am. J. Occup. Ther. 78, 7802180240 (2024).

    Article  PubMed  PubMed Central  Google Scholar 

  4. Nocilli, M., d’Avella, A. & Berger, D. J. A novel approach to promote upper-limb motor recovery in stroke survivors using assistive myoelectric control and adaptive visual feedback in virtual reality. Front. Bioeng. Biotechnol. 13, 1628679 (2025).

    Article  PubMed  PubMed Central  Google Scholar 

  5. Ghazavi Dozin, S. M., Mohammad Rahimi, N. & Aminzadeh, R. Wii Fit-based biofeedback rehabilitation among post-stroke patients: a systematic review and meta-analysis of randomized controlled trial. Biol. Res. Nurs. 26, 5–20 (2024).

    Article  PubMed  Google Scholar 

  6. Lu, Z. et al. Personalized stroke rehabilitation via stratified interpretable modeling with wearable IMUs. IEEE Trans. Neural Syst. Rehabil. Eng. 33, 4325–4337 (2025).

    Article  PubMed  Google Scholar 

  7. Stinear, C. M., Smith, M. C. & Byblow, W. D. Prediction tools for stroke rehabilitation. Stroke 50, 3309–3316 (2019).

    Article  Google Scholar 

Download references

Acknowledgements

This work was supported by the Shanxi Province Chinese Medicine Science and Technology special research project (2024ZYY2A023) and the Integration of Medicine, Engineering, and Basic Medical Sciences Special Project of Shanxi Provincial Health Commission (2025YGYL026). The funders had no role in the conceptualisation, writing, or decision to submit this manuscript.

Author information

Authors and Affiliations

  1. Academy of Medical Sciences, Shanxi Medical University, Taiyuan, People’s Republic of China

    Ziwei Hou

  2. Department of Neurology, Affiliated Cardiovascular Hospital of Shanxi Medical University; Shanxi Key Laboratory of Heart Failure Precision Medicine, Shanxi Cardiovascular Hospital, Taiyuan, People’s Republic of China

    Chen Chen

Authors

  1. Ziwei Hou
  2. Chen Chen

Contributions

Ziwei Hou: writing-original draft, visualisation. Chen Chen: conceptualisation, writing - original draft, writing-review & editing.

Corresponding author

Correspondence to Chen Chen.

Ethics declarations

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Hou, Z., Chen, C. Expanding the functional scope of AI-driven rehabilitation exergames: multimodal biofeedback and adaptive safety monitoring. npj Digit. Med. 9, 780 (2026). https://doi.org/10.1038/s41746-026-03141-z

Download citation

  • Received: 22 February 2026

  • Accepted: 05 August 2026

  • Published: 09 October 2026

  • Version of record: 09 October 2026

  • DOI: https://doi.org/10.1038/s41746-026-03141-z

来源:Nature 健康护理主题 · nature.com