根据研究统计，有近75%的女性有不同程度的骨盆前倾，随之而来的是腰、髋、膝关节损伤以及排尿、排便、分娩甚至呼吸方面等疾病发作的风险，需要通过康复运动帮助患者快速恢复到解剖学的最佳状态和最佳的生理功能。由于骨盆前倾康复运动周期长和康复医疗资源紧张的等现实情况，多数患者仅能选择居家自行完成康复运动。但是当前研究中尚未形成统一的康复运动评价标准，患者在自我康复运动过程中难以对自身动作规范性和达标程度实现正确评估，以及难以获得及时的个性化的康复运动分析和指导，无法保障居家康复的恢复效果和运动安全性。因此本研究聚焦于数字疗法中的“互联网+康复医疗”，对骨盆前倾康复运动智能指导系统进行了深入的理论和应用研究，主要工作及成果如下：1.制定康复运动评价标准：本研究通过两轮德尔菲法咨询，确定了具备专家共识和技术可实现性的康复运动质量评价模型，包括3个一级指标和7个二级指标，涵盖动作准确性、稳定性和有效性三个评价维度的指标方案，以及通过层次分析法所确定的各级指标权重。2.基于评价标准的智能评估：本研究采用BlazePose的人体骨骼关键点检测技术，提取14个人体关节角变化特征，并提出基于可移动窗口DTW周期截断机制的康复运动质量评价方法，实现了骨盆前倾康复运动质量评价指标的自动化计算。通过康复专家认可的骨盆前倾康复运动数据集验证，关节结构准确性指标计算的准确率在87%以上，时间指标除动作间歇时长外准确率均在89%以上。3.基于评估结果的智能指导：本研究使用自整理意图识别数据库、康复运动领域专业知识数据库和康复运动对话数据集，对传统大语言模型问答路径进行优化，实现康复运动智能问答。系统集成意图识别Agent，关联用户运动数据和专业知识，并对模型进行LoRA微调，在自动评估的各指标中高出ChatGPT4平均0.1027的绝对分差，并在人工GSB评估中较ChatGPT4实现59.2%的更优回复比例。4.构建远程数字康复平台：在上述成果基础上，设计了移动端的骨盆前倾康复运动智能指导系统，该系统集成运动跟练和智能问答两个核心模块，辅以用户信息收集和基础管理功能，为骨盆前倾患者提供了一个全面的自我康复运动的评估与干预平台。
According to research statistics, nearly 75% of women suffer from different degrees of pelvic tilt, resulting in injuries to the waist, hips, and knees, and diseases related to uri-nation, defecation, childbirth, and even breathing. Rehabilitation exercise is the main means of treatment to help patients recover. Due to a prolonged convalescence and the shortage of rehabilitation medical resources, most patients have to choose to do rehabilita-tion exercise at home. However, in the current home rehabilitation, there are not enough physiotherapists to instruct patients on proper exercise in their home, which leads chal-lenges in the effect and safety of rehabilitation. Therefore, aiming at creating an intelligent assistant for pelvic forward rehabilitation exercise, this study focuses on "Internet + reha-bilitation therapy" to do some work, including formulating a quality assessment criteria of rehabilitation exercise with expert consensus, assess the consistency of rehabilitation movements with standard intelligently by AI, and applying LLM to give personalized re-habilitation guidance and suggestions. The main work and achievements are as follows:1. Formulation of evaluation criteria for rehabilitation exercise: Through two rounds of Delphi consultation, this study formulated a quality assessment model of rehabilitation exercise with expert consensus and practicability, including 3 first-level indicators and 7 second-level indicators which are selected from three evaluation angles of action accuracy, stability and effectiveness. Then through the analytic hierarchy process, the weights of all levels of indicators were determined.2. Intelligent assessment based on evaluation criteria: In this study, BlazePose was used to detect the key points of human bones, based on which the sequential variation of 14 human joints is extracted. And then, on the basis of DTW with movable window, a method was proposed to identify the exercise repetitions and automatically calculate the quality assessment indicators of the pelvic forward rehabilitation exercise. In the self-built “pelvis-video” and “pelvis-image” dataset recognized by experts, the accuracy rate of joint structure accuracy index calculation is more than 87%, and the accuracy rate of time index is more than 89% except the interval time.3. Guidance with LLM based on assessment results: In this study, self-built intention recognition database, professional knowledge database and dialogue data set of rehabilita-tion exercise were made, an intention recognition agent was installed, user exercise data and professional knowledge were associated with RAG, and a general LLM was fine-tuned with LoRA, so that an intelligent Q&A system for pelvic forward rehabilitation ex-ercise was initially realized. The average score of the Q&A system in BLEU and ROUGE was 0.1027 points higher than that of ChatGPT4. In the manual GSB evaluation of 120 question responses, 59.2% of responses were considered better than ChatGPT4.4. Design of remote digital rehabilitation platform: On the basis of the above achievement, an intelligent guidance system for pelvic forward rehabilitation exercise at mobile terminal was designed. The system integrates two core modules of exercise train-ing and intelligent question and answer, and a basic functional module for user infor-mation collection and management, which provides a comprehensive self-rehabilitation exercise assessment and intervention platform for patients with pelvic forward.
