Fall risk detection method based on multi-modal health data

Abstract

The invention relates to the technical field of data prediction, in particular to a tumble risk detection method based on multi-modal health data. For the feature data of the user, tumble risk prediction is carried out through an XGboost method, and a first tumble risk probability corresponding to the current moment is obtained; for the feature data of the user, the prediction accuracy of the large model is improved through an SFT fine tuning mode, and a second fall risk probability corresponding to the current moment is obtained; fusing the first tumble risk probability and the second tumble risk probability to obtain a comprehensive tumble risk probability; and taking the comprehensive fall risk probability as a prior probability in the Bayesian theorem, and obtaining a fall prediction probability through the Bayesian theorem. According to the method, multi-modal fusion and dynamic Bayesian updating are realized, the sign time sequence and the static health information are subjected to fusion analysis for the first time, the model robustness is improved, individualized risk tracking is realized, and the method adapts to health changes.

Description

Technical Field

[0001] This invention relates to the field of data prediction technology, specifically to a fall risk detection method based on multimodal health data. Background Technology

[0002] With an aging population, falls have become one of the major health threats to the elderly, leading to fractures, brain injuries, and even death. Current fall detection technologies mainly include motion monitoring based on wearable sensors and environmental perception methods based on vision or radar. These methods are largely limited to single-modal data and have not yet fully integrated multi-source health information to form a dynamic risk model.

[0003] Existing fall risk prediction methods that rely on single-modal data tend to overlook individualized factors and lack dynamic tracking capabilities. Ignoring individualized factors includes: current technologies do not adequately consider the effects of medication or medical history, such as orthostatic hypotension caused by antihypertensive drugs or balance disorders due to diabetic neuropathy, leading to significant generalization bias in risk assessment; lacking dynamic tracking capabilities includes: single-point-of-time data cannot capture long-term changes in health status, and daily vital sign changes are not included in the model, making preventative early warning difficult. Summary of the Invention

[0004] To address the technical problem of low accuracy in preventative early warning systems, the present invention aims to provide a fall risk detection method based on multimodal health data. The specific technical solution adopted is as follows: Acquire user characteristic data, which includes: vital signs data, health data, and follow-up records; For user feature data, fall risk is predicted using the XGboost method to obtain the first fall risk probability at the current moment; for user feature data, the prediction accuracy of the large model is improved by SFT fine-tuning to obtain the second fall risk probability at the current moment; the first and second fall risk probabilities are combined to obtain the comprehensive fall risk probability. The combined fall risk probability is used as the prior probability in Bayes' theorem, and the fall prediction probability is obtained through Bayes' theorem.

[0005] Furthermore, the fusion of the first fall risk probability and the second fall risk probability to obtain the comprehensive fall risk probability includes: The first fall risk probability and the second fall risk probability are weighted and summed to obtain the comprehensive fall risk probability; wherein, the sum of the weight values ​​of the first fall risk probability and the second fall risk probability is 1.

[0006] Furthermore, the formula for calculating the overall fall risk probability is as follows: Prior_risk(t)=Prior_risk_xg(t)×N+Prior_risk_m(t)×(1-N); Where N is the weight value of the first fall risk probability; Prior_risk(t) is the comprehensive fall risk probability of falling at time t; Prior_risk_xg(t) is the first fall risk probability; and Prior_risk_m(t) is the second fall risk probability.

[0007] Furthermore, the phrase "using the comprehensive fall risk probability as the prior probability in Bayes' theorem" and the fall prediction probability obtained through Bayes' theorem include: Prior_risk(t∣t-1)=Prior_risk(t-1∣t)×Prior_risk(t) / Prior_risk(t-1); Where Prior_risk(t) is the overall probability of a fall occurring at time t; Prior_risk(t-1|t) is the probability of an actual fall occurring after the predicted fall risk; Prior_risk(t|t-1) is the predicted probability of a fall occurring at time t based on the actual probability of a fall occurring at time t-1; and Prior_risk(t-1) is the predicted probability of a fall occurring at time t-1.

[0008] Furthermore, the step of using the comprehensive fall risk probability as the prior probability in Bayes' theorem, and then obtaining the fall prediction probability through Bayes' theorem, further includes: A low-risk response is triggered when the predicted fall probability is greater than the preset low-risk threshold and less than or equal to the preset medium-risk threshold; a medium-risk response is triggered when the predicted fall probability is greater than the preset medium-risk threshold and less than or equal to the preset high-risk threshold; and a high-risk response is triggered when the predicted fall probability is greater than the preset high-risk threshold.

[0009] Furthermore, after obtaining the user's feature data, the process includes: A Kalman filter is applied to remove noise from the vital signs data; Based on the usage instructions of medication data in health data, drugs are classified. For disease data, the mapping is done as vector embedding.

[0010] Furthermore, a wrist-worn smart device is used to collect vital sign data at fixed time intervals, including blood pressure, heart rate, and pulse.

[0011] Furthermore, the mapping of disease data to vector embedding includes: One-Hot coding is used to represent disease categories, and different disease categories are weighted.

[0012] Furthermore, health data is obtained from the electronic medical record system through an API interface. This health data includes medication data, disease data, age, and vision.

[0013] Furthermore, the weighting value for the probability of the first fall is 0.7, and the weighting value for the probability of the second fall is 0.3.

[0014] The embodiments of the present invention have at least the following beneficial effects: This invention uses the XGboost method to predict fall risk based on user feature data, obtaining the first fall risk probability at the current moment. It then uses SFT fine-tuning to improve the prediction accuracy of the large model, obtaining the second fall risk probability at the current moment. The first and second fall risk probabilities are then fused to obtain a comprehensive fall risk probability. This comprehensive fall risk probability is used as the prior probability in Bayes' theorem, and the fall prediction probability is obtained through Bayes' theorem. This invention achieves multimodal fusion and dynamic Bayesian updating, and for the first time, it integrates and analyzes time-series vital signs with static health information, improving model robustness and enabling individualized risk tracking to adapt to changes in health. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating a fall risk detection method based on multimodal health data, provided in one embodiment of the present invention. Figure 2 This is a system block diagram of a fall risk detection system based on multimodal health data, provided as an embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a fall risk detection method based on multimodal health data proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] The following description, in conjunction with the accompanying drawings, details a specific scheme for a fall risk detection method based on multimodal health data provided by the present invention.

[0020] Example 1: Please see Figure 1 The diagram illustrates a flowchart of a fall risk detection method based on multimodal health data according to an embodiment of the present invention. The method includes the following steps: Step S100: Obtain the user's feature data, which includes: vital signs data, health data, and follow-up records.

[0021] User characteristic data includes: vital signs data, health data, and follow-up records.

[0022] The vital signs data include: blood pressure, heart rate, pulse wave, and other vital signs information collected using a wrist-worn smart device at three time periods: morning, noon, and evening. The wrist-worn smart device, if integrating a blood pressure sensor and a heart rate sensor, collects blood pressure data including systolic and diastolic blood pressure.

[0023] Health data includes: medication data, disease data, age, vision and other health information obtained from the Electronic Medical Record (EMR) system via API interface.

[0024] Follow-up records include: follow-up information on the user's gait, dizziness, fatigue, etc., collected by nursing staff.

[0025] After obtaining user feature data, the feature data is preprocessed: Vital data: Kalman filter was applied to remove noise; temporal features were extracted, including mean and variance.

[0026] Medication data: Based on the usage data, the medications are classified according to their instructions for use. The classified medication categories include diuretics, analgesics, antihypertensives, and psychotropic drugs.

[0027] Disease data: mapped to vector embeddings, such as using One-Hot encoding to represent disease categories and weighting their impact on falls, for example, Parkinson's disease has a weight of 0.4.

[0028] Step S200: For the user's feature data, fall risk prediction is performed using the XGboost method to obtain the first fall risk probability at the current moment; for the user's feature data, the prediction accuracy of the large model is improved by SFT fine-tuning to obtain the second fall risk probability at the current moment; the first fall risk probability and the second fall risk probability are fused to obtain the comprehensive fall risk probability.

[0029] Fall risk prediction based on XGboost: For the extracted user feature data, fall risk prediction is performed by using the traditional XGboost method to obtain the first fall risk probability Prior_risk_xg(t) corresponding to the current time t. The specific calculation process is as follows: Assume the training set is ,in For feature vectors, The labels are used. The model's prediction function is: Where F represents the function space of the regression trees; K is the number of regression trees.

[0030] The objective function consists of two parts: in, This represents the loss function, such as logistic loss; Ω is the regularization term. This is used to penalize complex tree structures, where T is the number of leaves, w is the weight of the leaf nodes, and n is the training set. The amount of data in the middle.

[0031] In this invention, the fall risk probability Prior_risk_xg(t) at the current time t is output by the trained XGBoost classification model, and its calculation formula is as follows: Where σ represents the logistic regression function sigmoid, which is used to map the model output to probability values.

[0032] It should be noted that the principle of XGBoost is well known to those skilled in the art and will not be elaborated here.

[0033] Fall risk prediction based on large model fine-tuning: For the extracted user feature data, a supervised fine-tuning algorithm (SFT) is used to fine-tune the large model to improve the prediction accuracy and obtain the second fall risk probability Prior_risk_m(t) corresponding to the current time t.

[0034] In this invention, Prior_risk_m(t) represents the predicted probability of a user's fall risk at the current time t after fine-tuning the pre-trained large model using SFT. This method is used to enhance the model's ability to understand complex health features and improve prediction accuracy.

[0035] Feature definition: Includes vital signs data (e.g., blood pressure, heart rate), health data (e.g., medical history, medication information, age, vision), and follow-up records (e.g., gait, dizziness, fatigue, etc.). Historical fall event records are labeled with binary classification labels (fall / not fall, 0 / 1). Each sample represents the user's state at a certain moment and the corresponding fall result, forming a supervised learning dataset. .

[0036] Loss function: in, This represents the probability of falling as output by the model.

[0037] During the inference phase, the user characteristics at the current time t are... Input the finely tuned model and output its corresponding fall risk probability: The output value ranges from [0,1], representing the probability that the user will fall at that moment.

[0038] It should be noted that the supervised fine-tuning algorithm (SFT) is a well-known technique in the art and will not be described in detail here.

[0039] By combining the first fall risk probability and the second fall risk probability, a comprehensive fall risk probability is obtained. More specifically, the first fall risk probability and the second fall risk probability are weighted and summed to obtain the comprehensive fall risk probability; wherein the sum of the weight values ​​of the first fall risk probability and the second fall risk probability is 1.

[0040] In one embodiment of the present invention, the weight value of the first fall risk probability is 0.7, and the weight value of the second fall risk probability is 0.3.

[0041] The formula for calculating the overall fall risk probability Prior_risk(t) is as follows: Prior_risk(t) = Prior_risk_xg(t) × N + Prior_risk_m(t) × (1-N); where N is the weight value of the probability of the first fall.

[0042] Step S300: The comprehensive fall risk probability is used as the prior probability in Bayes' theorem, and the fall prediction probability is obtained through Bayes' theorem.

[0043] Personalized calibration: Update user baseline risk using Bayesian inference: Prior_risk_t=Posterior_risk_{t-1}, and calculate Posterior_risk_t by combining new follow-up data.

[0044] Bayesian inference, based on Bayes' theorem, dynamically updates the early warning estimate of fall risk by combining prior probabilities and new evidence.

[0045] The probability of falling is taken as the prior probability in Bayes' theorem, and the fall prediction probability is obtained through Bayes' theorem.

[0046] The formula for calculating the fall prediction probability is: Prior_risk(t∣t-1)=Prior_risk(t-1∣t)×Prior_risk(t) / Prior_risk(t-1); Where Prior_risk(t) is the overall probability of a fall occurring at time t; Prior_risk(t-1|t) is the probability of an actual fall occurring after the predicted risk of a fall has occurred; Prior_risk(t|t-1) is the predicted probability of a fall occurring at time t based on the actual probability of a fall occurring at time t-1.

[0047] In a preferred embodiment of the present invention, after obtaining the fall prediction probability through Bayes' theorem, different levels of risk response are triggered based on the magnitude of the fall prediction probability: A low-risk response is triggered when the predicted fall probability is greater than a preset low-risk threshold and less than or equal to a preset medium-risk threshold; a medium-risk response is triggered when the predicted fall probability is greater than a preset medium-risk threshold and less than or equal to a preset high-risk threshold; and a high-risk response is triggered when the predicted fall probability is greater than a preset high-risk threshold. In this embodiment of the invention, the preset low-risk threshold is 0.65, the preset medium-risk threshold is 0.80, and the preset high-risk threshold is 0.95.

[0048] Feedback loop: Users label false positives for model fine-tuning and dynamic updates.

[0049] As a specific example of this invention, in a senior living community setting, users connect a fitness tracker to a mobile app to obtain vital sign data. They then manually input or voice-report follow-up records weekly via the app. The mobile app synchronizes data from the community health platform through a WeChat mini-program interface. For example, when a user is walking, the fitness tracker monitors a stable body temperature of 37°C, but the pulse is irregular (e.g., variability of 15%), and blood pressure fluctuates (e.g., diastolic blood pressure from 70 mmHg to 85 mmHg), with a peak heart rate of 85 bpm. Health data includes medication and disease data, such as calcium supplementation and osteoporosis ICD-10: M81. This week's follow-up record score is 28 points, with a trend showing a 5-point increase from last month, and the symptom being joint stiffness. The data is preprocessed; the mean and variance of the extracted vital sign data are calculated. The extracted follow-up and health data are then normalized. Then, a fusion model was used: the prediction result based on XGboost was Prior_risk_xg=0.62, and the fall risk prediction based on fine-tuning of the large model was Prior_risk_m=0.89. The fall prediction result for this time was 0.701. The previous prediction result was 0.80, and the Bayesian update baseline was from 0.55 to 0.60, resulting in a final Risk=0.80. Finally, an early warning response was implemented: the app pushed prevention suggestions such as "strengthen calcium intake and pay attention to gait balance," and the trend chart was reported to the community doctor. If the user feedback was "effective," it was used to strengthen the positive samples of the model.

[0050] Example 2: Please see Figure 2 The diagram illustrates a system block diagram of a fall risk detection system based on multimodal health data according to an embodiment of the present invention. The system includes the following modules: The feature extraction module is used to acquire the user's feature data, which includes: vital signs data, health data, and follow-up records. The fall risk analysis module is used to predict fall risk based on user feature data using the XGboost method, obtaining the first fall risk probability at the current moment; it also uses SFT fine-tuning to improve the prediction accuracy of the large model based on user feature data, obtaining the second fall risk probability at the current moment; and finally, it combines the first and second fall risk probabilities to obtain the comprehensive fall risk probability. The fall risk prediction module is used to take the comprehensive fall risk probability as the prior probability in Bayes' theorem and obtain the fall prediction probability through Bayes' theorem.

[0051] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0052] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A fall risk detection method based on multi-modal health data, characterized in that, The method comprises the following steps: obtaining feature data of a user, the feature data comprising: vital sign data, health data, and follow-up records; for the feature data of the user, performing fall risk prediction by an XGboost method to obtain a first fall risk probability corresponding to a current time; for the feature data of the user, improving the prediction accuracy of a large model by an SFT fine-tuning method to obtain a second fall risk probability corresponding to the current time; fusing the first fall risk probability and the second fall risk probability to obtain a comprehensive fall risk probability; using the comprehensive fall risk probability as a prior probability in Bayes' theorem to obtain a fall prediction probability by Bayes' theorem.

2. The fall risk detection method based on multi-modal health data according to claim 1, characterized in that, The fusing of the first fall risk probability and the second fall risk probability to obtain a comprehensive fall risk probability comprises: performing weighted summation on the first fall risk probability and the second fall risk probability to obtain a comprehensive fall risk probability; wherein the sum of the weight values of the first fall risk probability and the second fall risk probability is 1.

3. The fall risk detection method based on multi-modal health data according to claim 1, characterized in that, The calculation formula of the comprehensive fall risk probability is: Prior_risk(t) = Prior_risk_xg(t) x N + Prior_risk_m(t) x (1-N); wherein N is the weight value of the first fall risk probability; Prior_risk(t) is a comprehensive fall risk probability of fall occurrence at time t; Prior_risk_xg(t) is the first fall risk probability; and Prior_risk_m(t) is the second fall risk probability.

4. The fall risk detection method based on multi-modal health data according to claim 1, characterized in that, The using of the comprehensive fall risk probability as a prior probability in Bayes' theorem to obtain a fall prediction probability by Bayes' theorem comprises: Prior_risk(t|t-1) = Prior_risk(t-1|t) x Prior_risk(t) / Prior_risk(t-1); wherein Prior_risk(t) is a comprehensive fall risk probability of fall occurrence at time t; Prior_risk(t-1|t) is a probability of actual fall occurrence after predicting fall risk; Prior_risk(t|t-1) is a fall prediction probability of predicting fall at time t based on a probability of actual fall occurrence at time t-1; and Prior_risk(t-1) is a fall prediction probability of fall occurrence at time t-1.

5. The fall risk detection method based on multi-modal health data according to claim 1, wherein, After the using of the comprehensive fall risk probability as a prior probability in Bayes' theorem to obtain a fall prediction probability by Bayes' theorem, the method further comprises: when the fall prediction probability is greater than a preset low-risk threshold and less than or equal to a preset medium-risk threshold, triggering a low-risk response; when the fall prediction probability is greater than the preset medium-risk threshold and less than or equal to a preset high-risk threshold, triggering a medium-risk response; and when the fall prediction probability is greater than the preset high-risk threshold, triggering a high-risk response.

6. The fall risk detection method based on multi-modal health data according to claim 1, wherein, After the obtaining of the feature data of the user, the method comprises: applying a Kalman filter to remove noise from the vital sign data; classifying drugs based on the usage instructions of medication data in the health data; and For disease data, mapping into vector embedding.

7. The fall risk detection method based on multi-modal health data according to claim 1, wherein, Using a wrist smart device to collect physical data in a fixed time period, the physical data including: blood pressure, heart rate, pulse.

8. The fall risk detection method based on multi-modal health data according to claim 6, characterized in that, The mapping of the disease data into a vector embedding comprises: Using One-Hot encoding to represent disease categories and weighting different disease categories. 9.The fall risk detection method based on multi-modal health data of claim 1, wherein, Obtaining health data from an electronic medical record system through an API interface, the health data including: medication data, disease data, age, vision.

10. The fall risk detection method based on multi-modal health data according to claim 2, wherein, The weight value of the first fall risk probability is 0.7, and the weight value of the second fall risk probability is 0.3.

Images