衰弱被定义为一种因生理储备减少而对压力源抵抗力下降的综合征，这种状态是由多个生理系统的累积效应导致的，导致个体最终不良后果产生的可能性增加。衰弱状态的评估与干预对于心脏外科患者具有重大意义，主要包括以下几方面原因：（1）衰弱状态在心脏手术患者中广泛存在，术前约有4%患者处于衰弱状态，特别是高龄人群，衰弱比例可进一步增加到40%。而心脏外科术后患者衰弱发生率可超过20%。（2）衰弱状态可用于预测心脏手术围术期及远期不良事件及生活质量，既往已有多项研究证实了衰弱对高龄心脏手术患者围术期及远期死亡、不良事件预测效能的提升。由于能够更直观地反应患者抵抗手术打击的生理储备，衰弱状态已被认可成为高龄健康状况评估的关键因素，应在临床实践中增加对该信息的采集和关注。2018年欧洲心脏病学会年会公布的心肌血运重建指南中提到，现有的手术风险评分缺乏衰弱这一重要变量。2021年美国心脏病发表的冠状动脉血运重建指南同样推荐将衰弱评估纳入现有手术风险评分。（3）衰弱是患者术后恢复过程中常见的一种主观不良感受与症状负担，随着心脏外科技术的发展，手术死亡率及并发症发生率逐渐降低，患者术后的主观感受需要受到更多关注。（4）不同患者间的围术期衰弱状态变化可能存在差异，并且与远期不良预后以及生活质量存在相关性，对其的评估同样值得人们关注与重视。

虽然衰弱是心脏手术患者结局预测和健康状况评估的重要指标，但其有效性和临床应用受到难以评估和量化的限制。衰弱状态的评价主要包括活动能力、精神认知和营养状况等几个方面。既往各研究团队开发了多项衰弱评估量表，但普遍存在以下几方面问题，导致针对常规临床实践中最优衰弱评估工具的选择尚未形成共识：（1）既往衰弱量表普遍基于高龄人群开发，其在整体心外科手术患者人群中的评估效果及不良结局的预测效能尚不清楚，有待进一步验证。（2）现有衰弱量表普遍仅支持术前单一时间点的衰弱评估，对于手术后的患者，急性手术打击可导致患者生理储备大量消耗，从而进入暂时或长期的衰弱状态，大大增加了衰弱状态信息采集以及长期监测的难度，难以实现围术期衰弱状态的连续监测。（3）现有衰弱量表普遍存在评估繁琐或包含维度不全的问题，极大地影响了它们在真实世界中的临床应用。

通过患者自报的采集形式可能是更易于心脏外科手术患者接受、采集成功率更高的方式，尤其是对于难以完成物理测量的心脏手术术后患者。患者报告结局（patient-reported outcome，PRO）是指患者直接报告的关于自身健康状况、功能状态和治疗感受的反馈，对其进行准确评估与及时干预逐渐成为学术热点，部分PRO的改善与预后获益存在显著相关性。基于患者报告的量表采集形式可以便捷地获取连续、动态的衰弱相关健康信息，为术后衰弱状态的评价以及变化趋势分析所需的信息收集提供了新的途径。

另一方面，随着数据科学以及人工智能技术的发展，多模态非接触式监测信息被发现可能全面反映衰弱状态，可以用于术前简便、高效的预测心脏手术预后。例如，通过步态分析技术（步态周期、运动学、时空参数、动力学等）及最新的人工智能算法建模，可能预测到衰弱相关的所有关键因素，包括活动能力、营养状况、精神认知及情绪。前期研究已提示，仅仅增加步态中的姿势信息，即可显著提升心脏手术不良事件的预测效果。然而，目前尚无研究尝试应用人工智能方式全面汇总多模态非接触式监测信息形成衰弱评价指标，并进一步预测心脏手术预后。

基于上述问题，本研究在心脏外科手术人群中采集了传统衰弱量表、基于患者报告的衰弱评估条目信息以及步态视频为主的非接触式监测信息，旨在评估现有传统衰弱量表在心脏外科人群中的采集和预测效果，并分别针对不同临床应用场景，基于新技术或新方法，开发并验证两种新型衰弱状态评估工具。在研究的第一部分，我们同时采集了目前临床常用的5种传统衰弱量表，旨在探索传统衰弱量表对心外科手术患者的围术期衰弱测量性能，以及患者术前衰弱状态对围术期不良事件的预测效果；第二部分，我们开发了一种基于患者报告结局的围术期衰弱评价量表，并利用该量表结果，描述患者不同类型的衰弱变化轨迹，识别不良状态，并比较不同变化轨迹的远期预后差异；第三部分，我们建立了一种基于多模态非接触式监测信息分析的衰弱自动化评估工具，并探索了衰弱自动化评估工具产生的指标“AI衰弱指数”与心脏外科患者围术期及远期预后的相关性。

主要的研究内容及结果如下：

第一部分 传统衰弱量表对心脏外科手术患者衰弱测量性能及结局预测效果评价研究

背景

心脏外科手术患者的全面术前评估，以确定手术干预的风险和收益，对于指导患者临床决策至关重要。衰弱状态可用于预测心脏手术围术期及远期不良事件及生活质量，逐渐成为心脏手术患者评估与干预的新热点。由于能够更直观地反应患者抵抗手术打击的生理储备，衰弱状态已被认可成为高龄健康状况评估的关键因素，应在临床实践中增加对该信息的采集和关注。多项指南推荐将衰弱评估纳入现有手术风险评分和常规临床实践。但其有效性和临床应用受到难以评估和量化的限制。既往各研究团队开发了多项衰弱评估量表，但这些评估工具普遍存在评估繁琐、包含维度不全的问题，极大地影响了它们在真实世界中的临床应用。针对常规临床实践中最优评估工具的选择尚未形成共识。因此，本研究旨在探索目前临床常用的5种传统衰弱量表对心外科手术患者的围术期衰弱测量性能，以及患者术前衰弱状态对围术期不良事件的预测效果。

方法

本研究入选行择期心脏外科手术的患者，在手术前和出院前1天或出院当天分别采集5种衰弱量表，包括简易体能状况量表（Short physical performance battery, SPPB）、弗里德量表（Fried scale, Fried）、FRAIL量表、埃德蒙顿衰弱评估量表（Edmonton Frailty Scale, EFS）和临床衰弱量表（Clinical Frailty Scale, CFS）。研究将根据衰弱量表各自定义的衰弱判断标准，比较5种衰弱量表测得的衰弱发生率，各节点的采集成功率，以及术前衰弱评分对围术期不良事件的预测效能。

结果

本研究最终纳入547名符合研究条件的患者（研究人群整体平均年龄为52.2岁，标准差13.2；女性229人，占比41.9%）。测得的总体衰弱发生率分别为：SPPB（术前5.5%，术后24.0%，总体6.4%），Fried（术前2.9%，术后18.4%，总体5.1%），FRAIL（术前0.7%，术后17.1%，总体8.3%），EFS（术前9.1%，术后69.6%，总体17.2%），CFS（术前5.3%，术后95.4%，总体47.2%）。各量表采集成功率分别为：SPPB（术前96.2%，术后4.6%），Fried（术前95.2%，术后15.9%），FRAIL（术前99.6%，术后86.7%），EFS（术前94.0%，术后14.4%），CFS（术前99.6%，术后86.7%）。基于术前衰弱评分建立的围术期不良事件预测模型的受试者工作特征曲线下面积（Area under the receiver-operating characteristic curve, AUC）及95%置信区间（Confidence interval, CI）分别为：SPPB（AUC 0.562，95% CI 0.483-0.641），Fried（AUC 0.511，95% CI 0.430-0.591），FRAIL（AUC 0.530，95% CI 0.453-0.607），EFS（AUC 0.543，95% CI 0.459-0.627），CFS（AUC 0.512，95% CI 0.430-0.594）。而综合了5种衰弱量表信息的混合模型的AUC值为0.703（95% CI 0.634-0.772）。

结论

现有5种传统衰弱量表均无法单独应用于心外科手术患者衰弱状态评价与手术风险预测。各量表之间测得的衰弱发生率差异明显，量表采集成功率较低，尤其是对于出院前的采集。未来需要针对心外科手术患者，开发评价维度更加全面、操作更加简便的衰弱评价工具。

第二部分 基于患者报告结局的心脏手术患者围术期衰弱评价量表开发及应用研究

背景

心脏外科手术作为一种急性打击可导致患者生理储备大量消耗，患者术后衰弱现象十分普遍，在成年患者人群中可达20%以上，而且不同患者间的衰弱状态变化可能存在差异，并与远期不良预后以及生活质量相关。此外，衰弱也是患者术后恢复过程中的一种主观不良感受与症状负担，值得人们关注与重视。然而，鉴于采集的难度，很少有研究关注患者手术后的衰弱状态以及变化趋势。对于难以完成物理测量的心脏手术术后患者，通过患者自报的采集形式可能是更易于患者接受、采集成功率更高的方式。因此，本研究旨在开发一种基于患者报告结局的围术期衰弱评价量表，并利用该量表结果，描述患者不同类型的衰弱变化轨迹，识别不良状态，并比较不同变化轨迹的远期预后差异。

方法

本研究包含一项定性研究（队列1）、一项单中心、前瞻性研究（队列2）和一项多中心、前瞻性研究（队列3）。队列1入选择期心脏外科手术患者和心外科医务人员，通过半结构化访谈和专家咨询，形成初版衰弱量表。在队列2中进行量表优化并进行信效度验证。在队列2和队列3的择期心脏外科手术人群中应用新开发的衰弱量表，采集术后第1天至3个月的8个时间节点的数据，通过潜在类别分析区分不同类型的衰弱变化轨迹，并比较不同轨迹患者的预后差异。

结果

研究通过对31名心脏外科手术患者和9名医务人员的半结构化访谈，以及包含15名专家的德尔菲法专家咨询，形成了包含10项条目的初版衰弱量表。通过在纳入500名择期心脏手术患者的队列2中的优化与信效度验证，最终形成了包含5项条目的新版衰弱评估量表。队列3纳入了全国8家中心的2503名择期心脏手术患者，在队列2与队列3总计3003名患者中，2981名患者成功采集了至少3个节点数据。通过潜在类别分析方法，将术后衰弱变化轨迹被分成4类，类别1患者（523人）术后早期衰弱程度最重，且恢复速度慢，始终处于所有患者中衰弱程度最重的状态；类别2患者（963人）术后早期衰弱程度较重，但恢复速度快，术后3个月时基本已是衰弱程度最低的一类；类别3患者（585人）术后早期衰弱程度较轻，但在术后1周内衰弱程度呈持续加重状态，直到术后2周左右才开始减轻，术后3个月时衰弱状态仍较重；类别4患者（910人）衰弱程度从术后早期到术后3个月期间始终最轻。生存分析显示4种类别患者预后存在显著差异（Log-rank检验，p< 0.001），类别1患者术后3个月内死亡及非计划再入院发生率最高，类别3患者其次，类别2和类别4患者发生率接近，但均低于类别1和类别3患者。

结论

本研究开发并验证了一种基于患者报告结局的心脏手术患者围术期衰弱评价量表。该量表条目简单，易于采集，有望用于心脏外科手术患者早期及远期的衰弱状态持续监测，识别心脏外科患者术后衰弱变化的异质性，实现不良衰弱状态的早期预警与个性化管理。

第三部分 基于多模态非接触式监测的心脏手术患者衰弱监测工具研发及应用研究

背景

衰弱是预测心脏手术患者结局的重要指标，其概念已被大量临床医生、研究人员和政策制定者所接受，但其效能和临床应用受到难以评估和量化的限制。既往各研究团体开发了多项衰弱评估的风险评分，但这些工具应用所需的时间较长，不易于采集，极大影响了它们在真实世界中的临床应用。5米步行试验被开发用于简便的评估衰弱风险，并取得了肯定的预测效果。但该试验由于信息量有限，无法全面反映衰弱。仍需要开发能全面反映衰弱，且简便易用的评价工具。随着人工智能技术的发展，多模态非接触式监测信息被发现可能全面反映衰弱，可以用于简便、高效的预测心脏手术预后。通过步态分析技术（步态周期、运动学、时空参数、动力学等）及最新的人工智能算法建模，可能预测到衰弱相关的所有关键因素，包括活动能力、营养状况、精神认知及情绪，且操作简便。但是，目前尚无研究尝试应用人工智能方式全面汇总多模态非接触式监测信息形成衰弱评价指标，并进一步预测心脏手术预后。基于上述背景，本研究的主要目的为建立一种基于多模态非接触式监测信息分析的衰弱自动化评估工具。

方法

本研究为一项单中心、前瞻性队列研究，纳入择期心脏外科手术患者，分别于术前和出院前以非接触方式采集步态视频，形成传统衰弱量表结果与步态视频配对数据。研究将按4：1比例将数据随机分为建模组和验证组，基于人工智能算法，应用建模组数据建立衰弱自动化评估工具，并在验证组中测试对传统衰弱量表结果预测效能。同时探索衰弱自动化评估工具产生的指标“AI衰弱指数”对围术期及远期预后的预测效能。模型的评价指标为AUC。

结果

本研究最终纳入521名患者（平均年龄52.0岁，标准差13.3；女性218人，占比41.8%），共采集步态视频和传统衰弱量表配对数据592例。其中训练组471例，验证组121例。模型在验证组中对5种传统衰弱量表预测的AUC为在0.731至0.893之间。其中对于SPPB和FRAIL，衰弱自动化评估模型与医院衰弱风险评分（Hospital Frailty Risk Score, HFRS）预测效果无显著差异。对于Fried、EFS和CFS，衰弱自动化评估模型的预测效能显著优于HFRS模型。术前AI衰弱指数预测患者围术期不良事件模型的AUC值为0.710（95% CI 0.642-0.778），显著优于传统的手术风险评分欧洲心脏手术风险评分系统2（European System for Cardiac Operative Risk Evaluation Ⅱ, EuroSCOREⅡ）模型（0.710 vs 0.573, p=0.01），与美国胸外科医师协会（Society of Thoracic Surgeons, STS）模型（0.710 vs 0.672, p=0.11）和中国冠状动脉旁路移植手术风险评估系统2（SinoSCORE Ⅱ）模型（0.710 vs 0.647, p=0.34）无显著差异。术后AI衰弱指数对于远期不良事件预测模型的AUC值为0.754（95% CI 0.554-0.954）。

结论

本研究开发并验证了一种基于多模态非接触式监测的心脏手术患者衰弱监测工具，可以替代传统衰弱评价工具，方便快捷地实现衰弱自动化评估。该工具有望用于心外科患者术前衰弱状态评估，以指导手术决策。术后衰弱状态评估有望用于指导患者出院分流以及心脏康复。

Frailty has been defined as a syndrome of decreased resistance to stressors due to reduced physiological reserves. This state is the result of the cumulative effects of multiple physiological systems, and leads to an increased likelihood of an individual's ultimate adverse outcome. The assessment and intervention of frailty is of great significance for cardiac surgical patients for several reasons:(1) Frailty is widespread among cardiac surgical patients, with approximately 4% of patients in a frail state preoperatively, and the prevalence of frailty can further increase to 40%, especially in the elderly population. Moreover, the incidence of frailty in postoperative cardiac surgery patients can exceed 20%. (2) Frailty can be used to predict perioperative and long-term adverse events and quality of life in cardiac surgery, and several previous studies have confirmed the role of frailty in enhancing the prediction of perioperative and long-term mortality and adverse events in elderly cardiac surgery patients. The ability to assess a patient's physiological reserve and resistance to surgical stress makes frailty a pivotal factor in evaluating health status in advanced age. The collection and consideration of this information in clinical practice should be emphasized. The 2018 European Society of Cardiology guidelines on myocardial revascularization acknowledge the significance of frailty as a variable in existing surgical risk scores. The 2021 American Heart Association Guidelines for Coronary Artery Revascularization similarly advocate for the incorporation of frailty assessment into existing surgical risk scores. (3) Frailty is a prevalent subjective experience characterized by adverse feelings and symptoms in patients during postoperative recovery. With the advancement of cardiac surgical techniques and the subsequent decline in surgical mortality and complication rates, it is imperative to prioritize the subjective feelings of patients in the postoperative period. (4) Furthermore, perioperative frailty may vary among patients and is associated with poor long-term prognosis and quality of life, and its assessment also deserves attention.

Although frailty is an important indicator of outcome prediction and health status assessment in patients undergoing cardiac surgery, its validity and clinical application are limited by the difficulty of assessment and quantification. The assessment of frailty mainly includes several aspects such as mobility, cognitive function and nutritional status. Several frailty assessment scales have been developed by various research teams, but the following issues are common, resulting in a lack of consensus on the optimal choice of frailty assessment tools for routine clinical practice: (1) Previous frailty scales were generally developed based on the elderly population, and their efficacy and predictive effectiveness of adverse outcomes in the overall cardiac surgical patient population are unclear and need to be further validated; (2) Existing frailty scales generally only support the assessment of frailty at a single point in time before surgery. For postoperative patients, acute surgical stress results in significant depletion of physiological reserves, which can lead to a temporary or long-term state of frailty, greatly increasing the difficulty of collecting information on the state of frailty as well as long-term monitoring, and it is difficult to realize the continuous monitoring of the state of frailty in the perioperative period; (3) Existing frailty scales are generally cumbersome to administer or contain incomplete assessment domains, which greatly affects their clinical application in the real world.

A form of collection through patient self-report may be more acceptable to cardiac surgery patients and have a higher collection success rate, especially for postoperative crucial cardiac surgery patients who have difficulty completing physical measurements. Patient-reported outcome (PRO) is the feedback directly reported by patients on their health status, functional status, and perception of treatment, and its accurate assessment and timely intervention are becoming a hot topic in academia, with significant correlation between improvement in some PROs and prognostic benefit. The patient-reported scale-based collection format can conveniently obtain continuous and dynamic frailty-related health information, which offering a novel approach to collect the information required for the evaluation of postoperative frailty status and the analysis of change trends.

On the other hand, with advances in data science as well as artificial intelligence techniques, multimodal non-contact monitoring information has demonstrated potential to reflect the full range of frailty status and can be used to predict cardiac surgery prognosis easily and efficiently preoperatively. For example, modeling using gait analysis techniques (gait cycle, kinematics, spatio-temporal parameters, kinetics, etc.) and the latest artificial intelligence algorithms may predict all key factors associated with frailty, including mobility, nutritional status, cognitive function and mood status. Previous studies have suggested that simply adding postural information in gait can significantly improve the prediction of adverse events in cardiac surgery. However, no study has attempted to apply an artificial intelligence approach to comprehensively integrate multimodal non-contact monitoring information to form a frailty evaluation index and further predict cardiac surgery prognosis.

To address these research gaps, this study systematically collected three categories of frailty-related data in cardiac surgery populations: conventional frailty scale measurements, patient-reported frailty assessment entry-based information, and video-based gait analysis as a primary non-contact monitoring modality. This series of studies aim to evaluate the data collection feasibility and predictive validity of existing frailty instruments in cardiac surgical cohorts, while also developing and validating two innovative frailty assessment tools optimized for distinct clinical scenarios through technological and methodological innovations. In the first part, we concurrently implemented five clinically established frailty scales to assess their operational performance in perioperative frailty measurement and their prognostic utility for predicting adverse surgical outcomes. In the second part, we developed a PRO-based perioperative frailty assessment scale, utilizing longitudinal PRO data to characterize heterogeneous frailty trajectory patterns, identify high-risk states, and compare long-term outcome disparities across distinct trajectory subgroups. In the third part, we introduced an automated frailty assessment system leveraging multimodal non-contact monitoring data. This component examines the clinical relevance of the derived “AI-Frailty Index” in predicting both perioperative complications and long-term prognosis.

The main findings are as follow:

Part I. Evaluation of Traditional Frailty Scales for Frailty Measurement Performance and Outcome Prediction in Cardiac Surgery Patients

Background

Comprehensive preoperative assessment of cardiac surgery patients for determining the risks and benefits of surgical intervention is crucial for guiding clinical decision-making. Frailty has increasingly become a focal point in cardiac surgery evaluation and intervention, serving as a predictor of both perioperative and long-term adverse events as well as quality of life. Given its ability to directly assess patients' physiological reserve and tolerance to surgical stress, frailty has been recognized as a key factor in geriatric health assessments, warranting enhanced clinical attention and systematic documentation. Multiple guidelines recommend integrating frailty evaluation into existing surgical risk scores and routine clinical practice. However, its clinical utility remains limited by challenges in standardized assessment and quantification. Although various frailty assessment tools have been developed, their real-world application is hindered by operational complexity and incomplete multidimensional assessment, with no consensus on the optimal instrument for routine clinical use. Therefore, this study aims to investigate the perioperative frailty measurement performance of five commonly used traditional frailty scales in cardiac surgery patients and their predictive efficacy for preoperative frailty status on perioperative adverse events.

Methods

This study enrolled patients scheduled for elective cardiac surgery. Data collection included administration five frailty assessment instruments preoperatively and either one day prior to discharge or on the discharge day, including the Short Physical Performance Battery (SPPB), Fried Frailty Scale (Fried), FRAIL Scale, Edmonton Frailty Scale (EFS), and Clinical Frailty Scale (CFS). Using each scale’s predefined diagnostic criteria, we compared frailty prevalence rates, data completion rates across assessment timepoints, and the predictive accuracy of preoperative frailty scores for perioperative adverse events.

Results

The study ultimately included 547 eligible patients (mean age 52.2 years [standard deviation, SD 13.2]; 229 [41.9%] female). Frailty prevalence rates across assessment phases were as follows: SPPB (preoperative 5.5%, postoperative 24.0%, overall 6.4%), Fried (2.9%, 18.4%, 5.1%), FRAIL (0.7%, 17.1%, 8.3%), EFS (9.1%, 69.6%, 17.2%), and CFS (5.3%, 95.4%, 47.2%). Data completion rates for each scale were: SPPB (preoperative 96.2%, postoperative 4.6%), Fried (95.2%, 15.9%), FRAIL (99.6%, 86.7%), EFS (94.0%, 14.4%), and CFS (99.6%, 86.7%). For preoperative frailty scores predicting perioperative adverse events, the area under the receiver-operating characteristic curve (AUC) with 95% confidence intervals (CI) were as follows: SPPB (AUC 0.562, 95% CI 0.483–0.641), Fried (0.511, 95% CI 0.430–0.591), FRAIL (0.530, 95% CI 0.453–0.607), EFS (0.543, 95% CI 0.459–0.627), and CFS (0.512, 95% CI 0.430–0.594). The composite model integrating all five scales achieved an AUC of 0.703 (95% CI 0.634–0.772).

Conclusion

The five traditional frailty scales evaluated demonstrated limited clinical utility for standalone frailty assessment and surgical risk prediction in cardiac surgery patients. Substantial variations in frailty prevalence were observed across scales, accompanied by inconsistent completion rates, particularly during pre-discharge evaluations. These findings underscore the necessity to develop novel frailty assessment tools specifically designed for cardiac surgery populations, incorporating comprehensive multidimensional evaluation and streamlined implementation.

Part II. Development and Application of a Patient-Reported Outcome-Based Perioperative Frailty Assessment Scale for Cardiac Surgery Patients

Background

Cardiac surgery, as an acute physiological stressor, induces significant depletion of patients’ physiological reserves, leading to a high prevalence of postoperative frailty. Notably, over 20% of non-elderly patients develop postoperative frailty, with marked interindividual variability in frailty progression. This variability correlates with long-term adverse outcomes and quality of life deterioration. Furthermore, frailty manifests as a subjective adverse experience and symptom burden during postoperative recovery, warranting clinical attention. Nevertheless, few studies have systematically investigated postoperative frailty trajectories due to methodological challenges in longitudinal assessment. For postoperative patients unable to undergo physical measurements, patient-reported outcome (PRO) offers a patient-centered approach with higher compliance and data collection efficiency. This study aims to develop a PRO-based perioperative frailty assessment scale and utilize its outcomes to characterize heterogeneous frailty trajectory patterns, identify high-risk states, and compare long-term prognosis across distinct trajectory subtypes.

Methods

This investigation comprised three sequential components: a qualitative study (Cohort 1), a single-center prospective study (Cohort 2), and a multicenter prospective study (Cohort 3). Cohort 1 enrolled elective cardiac surgery patients and cardiac care professionals to develop a preliminary frailty scale through semi-structured interviews and expert panel reviews. Cohort 2 refined the scale and conducted reliability and validity testing. Cohort 3 implemented the optimized scale in elective cardiac surgery populations across multiple centers. Longitudinal data were collected at eight postoperative timepoints from day 1 to 3 months. Latent class analysis was employed to identify distinct frailty trajectory subtypes, with subsequent comparison of prognosis across trajectory patterns.

Results

Through semi-structured interviews with 31 cardiac surgery patients and 9 healthcare professionals, combined with Delphi expert consensus process involving 15 specialists, we developed a preliminary 10-item frailty scale. Subsequent refinement and psychometric validation in Cohort 2 (n=500 elective cardiac surgery patients) yielded an optimized 5-item frailty assessment scale. Cohort 3 enrolled 2,503 elective cardiac surgery patients across eight national centers. Among the combined 3,003 patients from Cohorts 2 and 3, 2,981 provided data from at least three time points. Latent class analysis identified four distinct postoperative frailty trajectory patterns. Class 1 (n=523): Severe initial frailty with slow recovery, maintaining the highest frailty burden throughout follow-up. Class 2 (n=963): Moderate initial frailty with rapid improvement, achieving the lowest frailty levels by 3 months. Class 3 (n=585): Mild initial frailty with progressive worsening peaking at 2 weeks, retaining substantial burden at 3 months. Class 4 (n=910): Consistently minimal frailty from immediate postoperative period to 3 months. Survival analysis revealed significant prognostic disparities among classes (log-rank test, p<0.001). Class 1 exhibited the highest 3-month mortality or unplanned readmission rates (14.7%), followed by Class 3 (9.2%), with Classes 2 (4.6%) and 4 (3.8%) demonstrating comparable lower risks.

Conclusion

This study successfully developed and validated a perioperative frailty assessment scale for cardiac surgery patients based on patient-reported outcome. Featuring concise items and efficient data collection, this scale enables continuous monitoring of frailty status during both early and long-term postoperative phases. It demonstrates strong potential for identifying heterogeneous progression patterns of postoperative frailty, facilitating early detection of high-risk states and guiding personalized intervention strategies.

Part III. Development and Application of a Multimodal Non-contact Monitoring Tool for Frailty Assessment in Cardiac Surgery Patients

Background

Frailty is a critical prognostic indicator for cardiac surgery outcomes, widely acknowledged by clinicians, researchers, and policymakers. However, its clinical utility remains constrained by challenges in assessment and quantification. Existing frailty risk assessment tools, though validated, are often time-consuming and operationally complex, limiting their real-world adoption. The 5-meter walk test was developed as a simplified frailty screening method with proven predictive value, yet its limited data capture fails to comprehensively characterize frailty. There is an unmet need for tools that holistically evaluate frailty while maintaining operational simplicity. Advancements in artificial intelligence (AI) have revealed the potential of multimodal non-contact monitoring to comprehensively evaluate frailty, enabling efficient prediction of surgical outcomes. Gait analysis techniques—including gait cycles, kinematics, spatiotemporal parameters, and kinetics—coupled with advanced AI modeling may capture all frailty-associated determinants such as mobility, nutritional status, cognition, and emotional well-being, with minimal procedural burden. Despite this promise, no studies have systematically integrated multimodal non-contact monitoring data into AI-driven frailty indices for prognostic prediction in cardiac surgery. Based on the above background, the main objective of this study is to establish an automated assessment tool for frailty based on the analysis of multimodal non-contact monitoring information.

Methods

This single-center prospective cohort study enrolled patients who underwent elective cardiac surgery. Gait videos were acquired via non-contact monitoring preoperatively and before discharge, paired with traditional frailty scale assessments. The dataset was randomly split into training and validation cohorts in a 4:1 ratio. We employed AI algorithms to develop an automated frailty assessment tool based on the training cohort data and evaluated its predictive accuracy for traditional frailty scale outcomes in the validation cohort. Additionally, we investigated the prognostic utility of the tool's output metric ("AI-Frailty Index") for perioperative and long-term clinical outcomes. Model performance was quantified using the area under the receiver-operating characteristic curve (AUC).

Results

The study ultimately included 521 patients (mean age 52.0 years [standard deviation, SD 13.3]; 218 [41.8%] females), yielding 592 paired gait video and traditional frailty scale datasets. The cohort was divided into a training set (n=471) and a validation set (n=121). The AI-based frailty assessment model achieved AUC values ranging from 0.731 to 0.893 in predicting the five traditional frailty scales within the validation cohort. For the SPPB and FRAIL scales, the AI model demonstrated comparable predictive performance to the Hospital Frailty Risk Score (HFRS). However, the AI model significantly outperformed HFRS in predicting Fried, EFS, and CFS scores. The preoperative AI-Frailty Index achieved an AUC of 0.710 (95% CI 0.642–0.778) for predicting perioperative adverse events, significantly higher than the European System for Cardiac Operative Risk Evaluation II (EuroSCORE II) model (0.710 vs. 0.573, p=0.01). No significant differences were observed compared to the Society of Thoracic Surgeons (STS) model (0.710 vs. 0.672, p=0.11) or the SinoSCORE II model (0.710 vs. 0.647, p=0.34). The postoperative AI-Frailty Index predicted long-term adverse events with an AUC of 0.754 (95% CI 0.554–0.954).

Conclusion

This study successfully developed and validated a multimodal non-contact monitoring tool for frailty assessment in cardiac surgery patients. The tool enables efficient, automated frailty evaluation and may potentially replace conventional frailty assessment instruments. Preoperative application of this tool may enhance surgical decision-making through streamlined frailty evaluation. Furthermore, postoperative frailty monitoring showed potential for guiding patient discharge planning and optimizing cardiac rehabilitation strategies.