Intelligent fall detection device

Abstract

This invention discloses an intelligent fall monitoring device, belonging to the field of fall monitoring technology, aiming to solve the problem of accuracy in fall monitoring and prediction for the elderly. It employs a collaborative artificial intelligence model architecture, integrating two different-dimensional artificial intelligence models for risk assessment. Model AI1, based on fuzzy logic, classifies fall risk levels into five levels: normal, low, medium, high, and emergency by real-time monitoring of vital signs such as heart rate, blood pressure, and blood oxygen saturation. Model AI2, based on deep belief networks, predicts risk by analyzing daily living activity patterns such as sitting, standing, walking, running, and jumping. The prediction results of the two models are input into a meta-model for comprehensive decision-making, generating the final low, medium, and high risk level prediction. For real-time fall monitoring, this invention proposes a hybrid monitoring algorithm, combining rule-based triggering based on acceleration and posture angle thresholds with AI fine-grained classification to achieve low latency and high accuracy.

Description

Technical Field

[0001] This invention belongs to the field of fall detection, and specifically relates to an intelligent fall detection device. Background Technology

[0002] Currently, falls are the leading cause of pain, disability, functional loss, and unnatural death among people aged 65 and older. A single fall can lead to a loss of confidence and create a vicious cycle of "fear of falling → reduced activity → declining physical function → higher risk of falls," severely impacting quality of life. Preventive activities can weaken muscle strength and balance, making people more prone to falls. However, continued physical activity after a fall is recommended because it can enhance strength, improve balance and coordination, and help regulate basic health indicators such as blood pressure, blood sugar, and weight. Therefore, prevention is crucial in reducing the incidence and severity of falls.

[0003] However, traditional risk assessment methods primarily rely on functional testing and clinical opinions. While helpful to some extent, these are largely post-hoc assessments, lacking effective predictive capabilities and requiring repeated testing. For complex fall events related to serious health conditions, existing methods lack universal applicability and dynamic adaptability, failing to meet the practical needs of continuous monitoring and early warning. Early AI models, while capable of predicting both simple and complex fall events, generally suffer from insufficient generalization ability, performance limitations imposed by specific datasets, and opaque (black box) algorithms, making them difficult to deploy in resource-constrained real-world monitoring equipment requiring reliable output. Solutions based solely on statistical calculations or single AI models often lack the ability to comprehensively process data from different sources and with varying properties, failing to effectively integrate multidimensional features such as physiological signs (e.g., blood pressure, heart rate, blood oxygen) and behavioral activities (sitting, standing, walking, running, jumping) for collaborative prediction.

[0004] The lack of real-time monitoring, tracking, and data transmission exacerbates the problem of fall risk management for elderly patients in healthcare institutions. In the vast majority of fall monitoring and prediction research, both the developed models and the research context lack the following standards: not all developed models can handle multiple profiles. Sometimes, pre-trained models are limited to understanding specific datasets. When changes occur, they fail to learn or provide output based on those changes. Most deep learning techniques used in system models for fall prediction are opaque, uncontrollable, and often ambiguous. These problems necessitate innovative approaches to designing healthcare monitoring systems that must operate with limited resources, such as system performance, and dynamically respond to fall events.

[0005] While combining IoT wearable devices and AI for assisted living (AL) and health monitoring (HM) is a hot research topic, current systems generally suffer from the following shortcomings: models typically rely on pre-defined rules or static models, lacking the ability to continuously learn and make adaptive decisions. Although deep learning technology offers high accuracy, its decision-making process lacks interpretability, and it is difficult to achieve real-time, low-power analysis on resource-constrained devices.

[0006] These issues highlight the need to design an innovative predictive model that is resource-efficient, low-latency, highly accurate, and capable of integrating multiple data sources to achieve truly early intervention. Summary of the Invention

[0007] While current fall monitoring technologies for the elderly are increasingly employing AI methods, they generally suffer from insufficient prediction accuracy, inflexible algorithms, high resource consumption, and difficulty in collaboratively processing multi-dimensional health data. This invention aims to overcome these shortcomings by proposing a collaborative AI model combining fuzzy logic and deep belief networks. Furthermore, it integrates the results of both models using a meta-model (random forest) to achieve more accurate, real-time, and reliable fall risk assessment and early prediction.

[0008] This invention aims to provide an intelligent fall monitoring device that, by integrating multi-dimensional data and performing multi-level collaborative decision-making, achieves accurate prediction of fall risk in the elderly and real-time, highly reliable monitoring of fall events. This invention effectively solves the problems of insufficient predictive ability, poor real-time performance, high false alarm rate, and high computational resource requirements in existing technologies.

[0009] The core of this invention lies in a collaborative dual-path intelligent system that performs two major functions in parallel: long-term fall risk prediction and real-time fall monitoring.

[0010] Long-term fall risk prediction path: It adopts an innovative collaborative artificial intelligence (AI) model architecture. It uses two independent AI sub-models (AI1 and AI2) to analyze the user's physiological signs and behavioral activity data respectively, and then uses a meta-model to make a comprehensive decision on the prediction results of the two to generate a unified long-term risk level.

[0011] Real-time fall detection path: A hybrid detection algorithm is adopted, which combines physical rule-based rapid threshold triggering with artificial intelligence-based fine behavior classification to achieve high accuracy and low latency in real-time fall detection.

[0012] These two pathways work together to form a comprehensive elderly health monitoring solution that emphasizes both "prediction" and "monitoring," and combines "offline risk assessment" with "real-time security alerts."

[0013] Multimodal data acquisition module:

[0014] Physiological data: Wearable devices (such as those with integrated sensors) monitor users' key vital signs in real time, such as blood pressure, heart rate, and blood oxygen saturation. These data serve as quantitative indicators of the body's internal state.

[0015] Behavioral activity data: Inertial measurement units (IMUs, such as accelerometers and gyroscopes) continuously collect data on users' activities of daily living (ADLs), including movement patterns and body posture information such as sitting, standing, walking, running, and jumping. This data reflects external activity capabilities and balance status.

[0016] Long-term risk prediction module: Collaborative AI model. This module is the core of predicting the future fall risk of the elderly.

[0017] The first sub-model (AI1) is a physiological risk prediction model based on fuzzy logic. This model receives real-time vital sign data and, through a pre-defined base of 111 fuzzy rules (which transform clear data into fuzzy linguistic variables and analyze them), combines expert knowledge and historical data to analyze physiological patterns, quantifying an individual's fall risk into five levels: normal (no risk), low, moderate, high, and urgent. The model emphasizes early detection of physiological abnormalities; for example, when blood pressure, heart rate, and blood oxygen levels show coordinated abnormal fluctuations, the model will raise the risk rating. Based on tests on publicly available datasets, the model achieves an accuracy of 95.24% and a sensitivity of 93.75%.

[0018] The second sub-model (AI2) is a risk prediction model for ADLs based on deep belief networks. This model receives pre-processed and feature-extracted ADL data. Its core is a deep belief network (DBN) composed of three layers of restricted Boltzmann machines (e.g., 256, 200, and 150 hidden units), trained using a combination of unsupervised pre-training (e.g., contrastive divergence) and supervised fine-tuning (e.g., the Adam optimizer). Through deep learning of daily activity patterns, the AI2 model can identify high-risk behavioral characteristics such as motor skill deterioration and balance loss, and classify fall risk into low, medium, and high levels.

[0019] Metamodel (Ensemble Decision Maker): A fusion decision model based on random forests. The predictions (i.e., their respective risk levels or scores) of AI1 and AI2 models are input features into the metamodel. Instead of simply weighting the results of the two sub-models, the metamodel constructs an ensemble learning classifier using the random forest algorithm. This classifier automatically learns the non-linear relationship and inherent patterns between the outputs of the two sub-models to determine the final, comprehensive fall risk level. For example, when AI1 predicts "medium risk" and AI2 predicts "high risk," the metamodel will synthesize the data to arrive at the most probable final risk rating. This collaborative architecture effectively reduces the bias of a single model, achieving an overall prediction accuracy of 90.00%.

[0020] Real-time fall detection module: Hybrid detection algorithm

[0021] This module focuses on the identification of instantaneous fall events.

[0022] Step 1: Quick triggering based on rule-based thresholds.

[0023] Acceleration Anomaly Monitoring: The system continuously monitors IMU data, especially the acceleration component a of the body along the x-axis. x When a is detected x The first alarm is triggered when the threshold exceeds the preset upper fall threshold (UFT) or falls below the lower fall threshold (LFT). This usually corresponds to a violent impact or a state of weightlessness at the moment of the fall.

[0024] Anomaly verification of posture angle: When the acceleration condition is triggered, the system immediately checks the gyroscope data to analyze whether the body posture angle has changed abruptly and deviated from the normal safe range (e.g., the angle between the body and the vertical direction is too large). Only when both conditions of "acceleration exceeding the threshold" and "abnormal posture angle" are met simultaneously is the system identified as a "suspected fall event," thereby activating subsequent AI verification. This dual verification significantly reduces false alarms caused by daily activities (such as sitting down quickly or jumping).

[0025] Step 2: AI-based fine-grained behavior classification.

[0026] For each "suspected event" triggered by a rule, the system will extract complete multi-sensor data within a time window centered on that moment.

[0027] These data features are input into a specialized AI classification model for final judgment, distinguishing between "true falls" and "non-fall similar behaviors" (False).

[0028] This hybrid algorithm combines the advantages of low latency (fast response) of rule-based judgment with high accuracy (low false alarms) of AI classification, achieving high-performance monitoring on resource-constrained devices. The final monitoring accuracy can reach 97.9%, with an average response time of only 1.45 seconds.

[0029] Data processing and output module:

[0030] Data preprocessing: The raw data acquired by the sensor needs to undergo noise reduction filtering. This invention proposes an improved filtering algorithm that effectively handles noise and abrupt changes in the signal by establishing a historical data buffer and performing sample bias calculation and replacement, thereby improving the reliability of subsequent analysis.

[0031] Risk assessment output: The output of the long-term risk prediction module (low / medium / high risk level) can be visualized through a user interface (such as a mobile APP) to help users and managers understand long-term risk trends and achieve early warning and intervention.

[0032] Real-time alarm output: Once the real-time monitoring module determines that it is a "real fall", the system will immediately send an emergency alarm notification to the pre-set mobile application of caregivers or monitoring centers through IoT cloud services (such as Blynk), and automatically attach the user's GPS location information so as to provide timely assistance.

[0033] The present invention provides an intelligent fall monitoring device, characterized in that it comprises: a wearable data acquisition device, a data processing and transmission module, and a control unit;

[0034] The control unit includes:

[0035] A rule-based threshold-based rapid triggering layer monitors events when acceleration exceeds a preset upper threshold or falls below a preset lower threshold, and simultaneously, a body posture angle exceeds a safe range. When these conditions are met, an event flagged as a suspected fall triggers a background AI verification layer. This layer extracts sensor data from a preset time window before and after the event and inputs it into a pre-trained deep neural network classifier to distinguish between a real fall and similar daily activities. If the intelligent verification layer classifies the event as a fall, the system immediately activates an emergency response mechanism.

[0036] The AI ​​verification layer includes:

[0037] The AI1 model is a fall risk prediction model based on fuzzy logic, which analyzes users' vital signs data and outputs a first risk level or score.

[0038] The AI2 model is a fall risk prediction model based on deep belief networks. It is used to analyze users' daily life activity data and output a second risk level or score.

[0039] The meta-model is configured to receive the outputs of AI1 and AI2 models and construct an ensemble learning classifier using the random forest algorithm. This classifier automatically learns the non-linear relationships and intrinsic patterns between the outputs of the two sub-models to determine the final, comprehensive fall risk level.

[0040] Wearable data collection devices are used to continuously collect users' vital signs and daily activity data, including:

[0041] Physiological sensor used to periodically measure blood pressure, pulse pressure, heart rate, and blood oxygen saturation;

[0042] The motion sensor, including at least one three-axis accelerometer and gyroscope, is used to acquire three-dimensional acceleration and attitude angle data of the body in real time;

[0043] The data processing and transmission module is located in the wearable data acquisition device and is used to filter and preprocess the acquired raw data and wirelessly transmit the data to the computing platform.

[0044] A collaborative intelligent prediction and analysis platform, deployed in the cloud or on a local server, contains at least two collaborative artificial intelligence models and a meta-model.

[0045] The early warning response and user interaction module, based on the comprehensive fall risk prediction results or real-time fall monitoring results, sends corresponding level of early warning or alarm information to the user and / or their caregiver via a mobile application.

[0046] The formula for determining whether the acceleration exceeds a preset upper limit threshold is:

[0047] Fall Detection = { True, if a x > UFT or a x < LFT; False, otherwise}

[0048] a x The x-axis component of acceleration represents the real-time x-axis acceleration value collected from the inertial measurement unit. During a fall, the human body may experience weightlessness or a severe impact, leading to a x A mutation occurs; this variable is used to capture anomalous features of linear motion.

[0049] UFT is the upper limit threshold: a preset upper limit value for acceleration, when a x > When UFT occurs, it indicates that high-speed movement or a strong impact may have occurred;

[0050] LFT is the lower threshold: a preset lower limit value for acceleration, when a x When the LFT is less than 1, it usually corresponds to a state of weightlessness;

[0051] If the acceleration condition is met, the algorithm will not immediately issue an alarm, but will instead initiate a second verification process. At this point, the gyroscope data is used to analyze the user's body posture angle.

[0052] After a fall, a person's posture changes from upright or sitting to an abnormal posture. The algorithm checks whether the user's posture angle exceeds the safe range;

[0053] The system only determines a fall event when both conditions are met: acceleration exceeding a threshold and abnormal attitude angle.

[0054] Real-time data acquisition and preprocessing: The sensor continuously acquires data, which is then filtered and denoised. The filtering algorithm includes:

[0055] In the filtering algorithm, the number of buffers N is set, X is the new sampled value; A is the set maximum sampling deviation value, the size of which needs to be judged based on experience; h is set as the minimum number of consecutive samples in the buffer whose deviation is less than A;

[0056] After filling the buffer, it is sorted. Each time a new sample value X is detected, the difference between the new sample value X and the historical sample values ​​in the buffer is calculated, where... For the i-th (i=1,2….N) sample value in the buffer, the maximum sample deviation is calculated as follows:

[0057]

[0058] Search Corresponding Recorded as Replace with the new sample value X This allows us to replace data that is not conducive to output criteria, and at the same time prevents valid signals from being judged as interference signals and discarded when the signal undergoes a sudden change.

[0059] Sort the data in the buffer again and retrieve it. (i=1,2….N-1) and Find the difference (j = i + 1, 2, ..., N). If the difference satisfies... and Then record ,in To achieve a number of consecutive samples with a deviation less than A, where , The sampled value is within the buffer.

[0060] Extract The maximum value of (k = 1, 2....N) , if < h, then Discard; if > h, then the corresponding sampling sequence , The average value of is output as

[0061]

[0062] where , is the sample number, and the result of this filtering is used as the final output value of the sampling.

[0063] The fall risk prediction method based on fuzzy logic implemented by the AI1 model includes the following steps:‌

[0064] Input step: Obtain the vital sign data of the user, including blood pressure, pulse pressure, heart rate and blood oxygen saturation;

[0065] Fuzzification step: Convert the clear numerical input into a linguistic variable through a trapezoidal membership function;

[0066] Inference step: Input the fuzzified data into a fuzzy rule base containing at least 111 preset rules for evaluation. The rules are used to classify the user's fall risk into five levels: normal, low, medium, high, and emergency;

[0067] Defuzzification step: Convert the fuzzy inference result into a clear risk score or grade label for output;

[0068] Output step: Use the predicted fall risk score or grade as the output of the AI1 model for further decision-making by the meta-model.

[0069] The fall risk prediction method based on the deep belief network implemented by the AI2 model includes the following steps:‌

[0070] Input processing step: Identify the user's daily activities based on the three-dimensional accelerometer data, including sitting, standing, walking, running, and jumping, and assign preset weights related to their importance to each activity. Combine the fall history record to form a feature vector;

[0071] Model construction step: Construct a deep belief network composed of multiple stacked restricted Boltzmann machine layers, perform layer-by-layer greedy pre-training, and use the Adam optimizer for supervised fine-tuning on the labeled data;

[0072] Risk classification steps: The network learns the intrinsic connection between user activity patterns and fall risk, and finally classifies users' fall risk into three levels: low, medium and high, and outputs the corresponding level or score;

[0073] Output steps: The predicted fall risk level or score is output as the AI2 model for further decision-making by the meta-model.

[0074] The meta-model-based collaborative fall risk prediction method includes the following steps:

[0075] Input mapping step: Receive the output of AI1 model (XAI1) and the output of AI2 model (XAI2), and optionally assign dynamic or static weights (WAI1 and WAI2) to each input.

[0076] Meta-classifier construction and training steps: Using the outputs of AI1 and AI2 models as input features, and Morse Fall Score (MFS) risk levels (low, medium, high) as labels, train a random forest classifier;

[0077] Prediction and decision-making steps: For new input features, prediction is performed using a trained random forest meta-model, and the final, unified fall risk level (Rpred) is determined by majority voting of all decision trees.

[0078] Output steps: Output a combined risk score (Scombined) and / or a final fall risk prediction rating (Rpred) for each user.

[0079] Artificial intelligence-based collaborative meta-model fall risk prediction algorithm:

[0080] Input mapping: Let: X AI1=AI1

[0081] XAI2=AI2

[0082] Assign weights WAI1 and WAI2 to the output:

[0083]

[0084]

[0085] in, :normal; :Low; :medium; :high; :urgent;

[0086] Comprehensive Risk Score: The comprehensive risk score Scombined is calculated as follows: Scombined = (WAI1 + WAI2) / 2;

[0087] Risk level definition: Risk level R is defined based on Scombined:

[0088]

[0089] Training a random forest meta-model includes training features Xtrain={WAI1,WAI2} • training labels; Ytrain = risk levels based on MFS { , , Random Forest Model: Initialize T, number of decision trees, n_estimators=100 max_depth=10 min_samples_split=4;

[0090] For each tree t in T:

[0091] Randomly select a subset Xt from Xtrain.

[0092] Decision trees are constructed using Xt and Ytrain.

[0093] The data is split at each node using features {WAI1, WAI2}.

[0094] The predictions from all T trees are aggregated through majority voting.

[0095] Predicting fall risk: For test data Xtest={WAI1,WAI2}

[0096] Each test instance is passed through all T decision trees.

[0097] The predicted risk level Rpred is calculated based on the majority vote of T trees.

[0098] For each participant i, the output is: participant i: {Scombined, Rpred};

[0099] Long-term, non-invasive fall risk assessment is conducted by continuously monitoring vital signs and activity patterns.

[0100] When a high risk of fall is predicted, an alert is issued to the user and their care network, prompting them to take preventative measures.

[0101] When a real-time fall event is detected, an emergency alarm is automatically triggered, sending an alert containing the user's location information.

[0102] This invention provides an intelligent fall detection device, which has the following beneficial effects:

[0103] Precise and comprehensive predictions: It innovatively combines fuzzy logic and deep belief networks to make synergistic predictions from both physiological (internal) and behavioral (external) dimensions, achieving a long-term risk prediction accuracy of up to 90.0%.

[0104] Fast monitoring and low false alarm rate: The two-stage hybrid monitoring algorithm of "rule triggering + AI verification" ensures a fast response within 1.45 seconds, while effectively filtering out non-fall activities through fine-grained AI classification, achieving a monitoring accuracy of 97.9%.

[0105] An improved filtering algorithm effectively handles noise and abrupt changes in signals by establishing a historical data buffer and performing sample bias calculation and replacement, thereby improving the reliability of subsequent analysis.

[0106] High resource efficiency and easy deployment: The efficient filtering at the rule layer avoids performing AI calculations on all data, enabling complex intelligent models to run stably on wearable devices with limited power consumption and computing power.

[0107] Highly practical and scalable: The system can be integrated into wearable devices for all-weather, unobtrusive monitoring. Its collaborative AI architecture also reserves space for incorporating more modal data (such as environmental information) in the future. Attached Figure Description

[0108] Figure 1 This invention relates to an intelligent fall detection device. Detailed Implementation

[0109] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0110] This invention introduces a novel collaborative AI model that predicts the future risk of falls in older adults based on behavioral and health status abnormalities. Accurate predictions are generated by combining the predictions of two AI models: AI1, based on vital signs and using fuzzy logic, and AI2, based on activities of daily living (ADLs) and using a deep belief network (DBN). A meta-model then combines these outputs to generate a total fall risk prediction.

[0111] Traditional fall risk assessments typically rely on functional tests and clinical opinions, which, while helpful, lack predictive power and require follow-up testing. This invention introduces a novel collaborative artificial intelligence model capable of predicting future falls in older adults with greater accuracy and effectiveness. Two distinct AI models are specifically combined with a meta-model to predict the probability of future falls in older adults. The model is first calibrated using data downloaded from a public repository. Utilizing multimodal datasets, such as physiological, behavioral, and functional assessments, the model aims to bypass the limitations of traditional classifiers and improve fall risk stratification. The established strategy has been initially validated using public datasets to ensure stability before incorporating real-time information in future phases. The proposed AI-driven solution has the potential to enable early intervention, reduce the likelihood of fall injuries, and empower older adults to live independently.

[0112] Please see Figure 1 The fall detection algorithm proposed in this invention adopts a hybrid architecture, which cleverly combines rule-based rapid response with artificial intelligence accurate classification to achieve the monitoring goal of high accuracy and low latency.

[0113] Rule-based threshold monitoring enables rapid preliminary judgment:

[0114] The algorithm first performs a rapid and computationally inefficient preliminary screening of sensor signals using a set of preset physical thresholds. This acts as the system's "first line of defense," designed to detect potential fall events as soon as possible.

[0115] When a person falls, their acceleration and posture undergo drastic and specific changes within a short period. The algorithm sets thresholds for these physical quantities, triggering a fall alarm when the monitored data exceeds the threshold.

[0116] The key to the algorithm lies in the setting of the threshold, which needs to be customized for the target population (such as elderly people of different weights and heights) to balance sensitivity and specificity and reduce false alarms.

[0117] The algorithm in this invention does not simply rely on a single acceleration threshold, but rather employs a multi-stage, multi-criteria decision-making process. This significantly improves the accuracy of monitoring.

[0118] Its core decision-making logic can be broken down into the following two key steps:

[0119] The system continuously monitors changes in the user's body acceleration using an accelerometer. Falls are usually accompanied by weightlessness and impact, which can cause a sharp change in acceleration that exceeds the range of normal activities (such as walking or sitting).

[0120] The algorithm first determines whether the acceleration exceeds a preset upper limit threshold. This is a preliminary, highly sensitive screening process used to detect all possible fall events.

[0121] Formula Fall Detection = { True, if a x > UFT or a x < LFT; False, otherwise} can be understood as the specific implementation of the first step above, "acceleration mutation monitoring," on the x-axis.

[0122] a x (Acceleration x-axis component): Represents the x-axis acceleration value acquired in real time from an inertial measurement unit (such as MPU-6050). During a fall, the human body may experience weightlessness (such as falling backward) or a violent impact (such as falling forward), resulting in a x A sudden change occurs. This variable is used to capture anomalous features of linear motion.

[0123] UFT (Upper Fall Threshold): The preset upper limit for acceleration. When a x > When UFT occurs, it indicates that a high-speed movement or a strong impact may have occurred (such as the impact at the moment of falling).

[0124] LFT (Lower Fall Threshold): A preset lower limit for acceleration. When a x When LFT is less than 1, it usually corresponds to a state of weightlessness (such as the free fall phase at the beginning of a fall).

[0125] If the acceleration condition is met, the algorithm will not immediately issue an alarm, but will instead initiate a second verification procedure. At this point, gyroscope data is used to analyze the user's body posture angles.

[0126] After a fall, a person's posture changes from upright or sitting to an abnormal posture such as lying down. The algorithm checks whether the user's posture angle exceeds the safe range (e.g., the angle with the vertical is too large).

[0127] The system only determines a fall event when both conditions, namely "acceleration exceeding the threshold" and "abnormal attitude angle", are met simultaneously.

[0128] The dual verification mechanism of acceleration and attitude angle effectively reduces the possibility of false alarms in daily activities.

[0129] The algorithm proposed in this invention is an ingeniously designed and clearly defined threshold monitoring scheme. By fusing information from multiple sensors, it greatly improves the specificity of fall detection while ensuring high sensitivity, making it an effective technical approach to solving the problem of safety monitoring for the elderly.

[0130] After the initial trigger, the system inputs a large amount of sensor data in a short period of time into the trained artificial intelligence model for more refined analysis in order to distinguish between real falls and similar daily activities (such as sitting down quickly or jumping), thereby significantly reducing the false alarm rate.

[0131] Artificial intelligence models, by learning from large amounts of labeled data (fall and non-fall scenarios), can capture more complex and subtle movement patterns, thus making more reliable judgments than simple rules.

[0132] The two do not work independently, but rather form a connected, collaborative decision-making pipeline, which can be summarized in the following steps:

[0133] 1. Real-time data acquisition and preprocessing: The sensor continuously acquires data, which is then filtered and noise-reduced.

[0134] This invention proposes a filtering algorithm based on the characteristic that when a person falls, their acceleration and posture undergo drastic and specific changes within a short period of time. The algorithm flow is as follows:

[0135] In the filtering algorithm, the number of buffers N is set, X is the new sampled value; A is the set maximum sampling deviation value, the size of which needs to be judged based on experience; h is set as the minimum number of consecutive samples in the buffer whose deviation is less than A.

[0136] After filling the buffer, it is sorted. Each time a new sample value X is detected, the difference between the new sample value X and the historical sample values ​​in the buffer is calculated, where... For the i-th (i=1,2….N) sample value in the buffer, the maximum sample deviation is calculated as follows:

[0137]

[0138] Search Corresponding Recorded as Replace with the new sample value X This allows data that is not conducive to output criteria to be replaced, while preventing valid signals from being judged as interference signals and discarded when the signal undergoes a sudden change.

[0139] Sort the data in the buffer again and retrieve it. (i=1,2….N-1) and Find the difference (j = i + 1, 2, ..., N). If the difference satisfies... And , record , where is the number of consecutive samplings to achieve a deviation less than A, where , is the sampling value in the buffer.

[0140] Take out the maximum value of (k = 1, 2....N) , if <h, then discard ; if >h, then output the average value of the corresponding sampling sequence , as

[0141]

[0142] where , is the sample serial number, and the result of this filtering is used as the final output value of the sampling.

[0143] The main advantages of this filtering algorithm are as follows:

[0144] (1) Set a historical sampling value buffer, and use the historical sampling value as the output criterion, which can improve the reliability of the filtering result.

[0145] (2) When performing amplitude limiting processing on the sampling sequence, use the average value of the sampling values corresponding to the maximum number of consecutive samplings as the output, which improves the anti-shake ability of the algorithm. It significantly improves the filtering effect on signals with large fluctuations.

[0146] (3) When processing signals with large jumps, the improved filtering algorithm uses the method of finding the sample with the largest deviation corresponding to and recording it as , and replacing with the new sampling value X. This ensures that each sampling value can enter the buffer and be used as the filtering output criterion, enabling the filtering output to have a smooth transition and improving the accuracy of the output.

[0147] 2. Fast triggering at the rule layer: Only when both conditions of "acceleration exceeds the threshold" and "attitude angle is abnormal" are satisfied, the artificial intelligence module is activated. This avoids sending all data into the model and greatly saves computing resources.

[0148] 3. Fine classification at the artificial intelligence layer: For each "suspected event", the system extracts the multi-sensor data features within a time window and inputs them into the intelligent model for classification.

[0149] 4. Final Decision and Response:

[0150] If the model classifies it as a "fall", the system immediately sends an alert to the caregiver's mobile application via the Blynk cloud service, along with GPS location information.

[0151] If the model classifies the event as "non-fall", the system will ignore the trigger and continue normal monitoring, thus effectively avoiding false alarms.

[0152] This combined strategy fully leverages the advantages of both methods:

[0153] High response speed: Rule-based threshold monitoring ensures that potential fall events can be detected instantly.

[0154] High accuracy and low false alarm rate: The intervention of artificial intelligence has greatly improved the ability to identify complex actions and effectively filtered out false alarms.

[0155] Resource efficiency: Preliminary screening through rules avoids complex model calculations on all data, making real-time intelligent monitoring possible on resource-constrained embedded devices.

[0156] Ultimately, the hybrid algorithm achieved an accuracy rate of 97.9% and an average response time of only 1.45 seconds, significantly improving the system's security capabilities for people with special needs.

[0157] This invention predicts future falls in older adults based on vital signs and activity level monitoring. Our meta-model is easily integrated with proposed collaborative AI models based on vital signs and activity level data, yielding highly accurate results that closely approximate real-world data obtained using a multi-fault scoring method. This proposed work extends our previous work, where the first model used only vital signs to predict future fall risk in older adults, and the second model used only activities of daily living (ADLs). Now, in this proposed work, we combine our two collaborative AI models, blending them with the meta-model and ground truth (i.e., the multi-functional gait sensor MFS), to produce a final fall risk prediction output. This novel approach enables our model to create timely alerts for fall predictions for older adults using real-time data.

[0158] This model is used to predict impending falls in older adults and categorizes fall risk levels for those aged 60 and older. Based on an assessment by an AI-driven collaborative meta-model, fall risk is classified into low, medium, and high levels. The predictive output is derived through collaboration between two AI models.

[0159] Fuzzy fall risk prediction system:

[0160] An innovative approach was employed to predict fall risk in older adults by using fuzzy logic analysis of vital signs such as heart rate, blood pressure, and blood oxygen saturation. Combining fall risk factors with vital sign monitoring is crucial for fall prediction. Vital signs are important indicators of an individual's overall health and current medical condition. They are among the most sensitive parameters of a healthy life, changing with individual lifestyle; therefore, regular monitoring is essential, especially in older adults, as any imbalance can lead to falls. Logically, the higher the frequency of vital sign measurements, the faster clinical deterioration can be noticed. The earliest evidence of potentially dangerous physiological changes or disturbances in the body can often be found in vital signs, which can also serve as the first indication that a disease has stabilized.

[0161] The data for this invention comes from the demo version 2.0.1 of the public warehouse eICU collaborative research database - PhysioNet - to verify the accuracy of the model. Data from the public warehouse is categorized into fallers and non-fallers, with each category sent separately to the fuzzy model. The fuzzy prediction model collects vital sign data such as blood pressure, pulse pressure, heart rate, and blood oxygen saturation. Using membership functions and fuzzy rules, the model classifies fall risk as normal, low, moderate, high, or urgent. This approach focuses on real-time identification of fall risk, aiming to mitigate fall-related injuries or deaths among older adults. The system collects data from three different sources, which are then processed and evaluated and validated according to the MFS (Fuzzy Fall Prediction Model). The proposed fuzzy-based fall prediction model (AI1 model) aims to identify older adults at risk of falls in order to minimize the economic and personal burden of fall-related injuries. This developed model will be a good companion for older adults, especially those living alone, as it educates them about fall risk and methods to prevent further falls. The model focuses on continuous monitoring of vital signs (blood pressure / heart rate / blood oxygen saturation, etc.) to detect fall-related anomalies early. Integrating this model into a smartwatch for real-time monitoring can alert older adults and their caregivers in most cases, enhancing proactive fall prevention measures. A key advantage of this approach is its unique use of vital signs for prediction, offering higher accuracy compared to the MFS (Multi-Fall Prediction System), categorizing fall risk into five levels: normal, low, moderate, high, and urgent. These findings underscore the effectiveness of the AI1 model in enhancing fall prediction strategies for older adults.

[0162] Fall risk assessment system based on deep belief networks:

[0163] This method identifies older adults at high risk of falls by continuously monitoring their activities of daily living (ADLs) and detecting anomalies. To demonstrate the ability of deep learning to predict early fall risk, the model is built on a deep belief network (DBN) and employs advanced AI techniques such as pre-training with contrastive divergence, fine-tuning with backpropagation, and minimizing the loss using the Adam optimizer. This could lead to more timely intervention and reduce the frequency and severity of falls in older adults.

[0164] The AI2 model monitors Activities of Daily Living (ADLs), including sitting, standing, walking, running, and jumping, and detects anomalies through continuous observation. The effectiveness of the proposed model is evaluated by comparing the predictions with traditional fall prediction techniques and ground truth (GT). Contemporary deep learning techniques can be effectively used to improve early fall risk prediction, thereby reducing the likelihood and severity of falls in older adults. Fall risk is categorized into low, medium, and high levels. An innovative model employs a Deep Belief Network (DBN) to predict falls in older adults using only ADLs and medical history. The weight of each ADL is determined based on its importance in fall prediction: sitting (0.2), standing (0.4), walking (0.6), running (0.8), and jumping (1.0). These weighted inputs are evaluated using a developed DBN-based fall risk prediction algorithm (DBN-FRPA). Pre-training and fine-tuning of the DBN model utilizes several Restricted Boltzmann Machine (RBM) architectures. The model demonstrated strong predictive power in assessments of the Morse Fall Scale (MFS), improving the prediction of falls among older adults and enhancing their safety.

[0165] AI-based collaborative fall risk assessment system:

[0166] Different types of comparators are used to combine the results of AI1 and AI2 to generate a predictive output on the future fall risk of older adults. Examples include weighted average or score fusion, rule-based fusion methods, and decision tree-based methods. However, all the fusion models discussed have significant drawbacks, such as their limited ability to predict the output of the training data. In the fall risk prediction meta-model system proposed in this invention, the output values ​​of AI1 and AI2 are used as input features to train a combined model, generating predictions in the meta-model-based fall risk prediction system. A random forest algorithm is used in the initial step to train the meta-model. The collaborative meta-model fall risk system algorithm is designed by combining the outputs of the AI1 and AI2 models.

[0167] To integrate the two artificial intelligence models, a comparator is used to effectively combine the results of AI1 (a fuzzy fall risk prediction model) and AI2 (a deep belief network-based fall risk prediction model), ultimately achieving improved predictions of future fall risk for older adults.

[0168] Meta-modeling strategies enable systems to learn from and update their decision models using past data, as opposed to rigid rule-based systems. This characteristic allows classifiers to improve their generalization ability, making their predictions more consistent and reducing the likelihood of false positives. Instead of directly weighting or averaging the predictions of AI1 and AI2, this approach extracts useful patterns and relationships between their predictions. The meta-classifier considers previous cases derived from these AI models to adaptively select the optimal weights for each provided AI model. The Random Forest algorithm was chosen as the primary training procedure for our meta-classifier due to its recognized efficiency and reliability in predictive analytics. Its main advantages are as follows:

[0169] It reveals the potential nonlinear and chaotic correlation between the outputs of AI1 and AI2, which could potentially improve prediction accuracy.

[0170] Unlike logistic regression, which assumes linearity, random forests are capable of handling complex decision boundaries.

[0171] It constructs many decision trees and averages the results, thereby eliminating overfitting.

[0172] This allows the model to be generalized to other datasets. Since both AI1 and AI2 can generate different risk scores or categories, the random forest algorithm automatically identifies the input features that significantly contribute to the final prediction.

[0173] Online learning of random forest meta-classifiers will create a data-driven and adaptive fall risk assessment model for older adults, outperforming existing fusion methods. The system applies ensemble learning principles to provide more accurate, interpretable, and scalable fall risk assessments for older adults.

[0174] A novel collaborative AI-based fall risk assessment model architecture is proposed:

[0175] This invention aims to predict the future fall risk of older adults based on the fusion of two key health indicators: vital signs and activities of daily living (ADLs), within a collaborative, AI-based fall risk assessment model. This model seeks to maximize predictive accuracy by integrating a range of artificial intelligence techniques, thereby contributing to timely and accurate fall risk estimation with high consistency with the Primary Functional Score (MFS), achieving an extremely high consistency rate of approximately 90%. The design comprises two independent AI models, one assessing vital signs and the other assessing ADLs, which are then fused into a meta-model through ensemble learning to enhance the overall predictive results. The first AI model (AI1) uses fuzzy logic to assess vital sign patterns, where 111 pre-defined fuzzy rules are used to categorize fall risk into five classes: normal (no risk), low, moderate, high, and urgent. If no abnormalities are detected, the model predicts normal, indicating no fall risk; while detected abnormalities are graded according to severity. The second AI model (AI2) uses deep belief networks (DBNs) to analyze ADL (Activities of Daily Movement) patterns and predicts fall risk based on learned historical data, classifying fall risk as low, moderate, or high. ADLs provide crucial information about mobility, balance, and movement patterns, enabling AI2 to identify individuals with impaired movement and increased fall risk. Following the predictions from both models, their outputs are fed into a meta-model, a decision model, to refine the final predictions.

[0176] The meta-model-based fall risk assessment system is a composite model that trains a meta-model classifier that uses the outputs of AI1 and AI2 as input features to generate consistent predictions. The collaborative meta-model fall risk prediction algorithm based on the final fall risk prediction is formulated by integrating the outputs of the AI1 and AI2 models. The meta-model determines the final risk level based on the combined input of AI1 and AI2, thereby reducing the inherent error in individual models and improving the reliability of the prediction. For example, if AI1 predicts "moderate fall risk" while AI2 predicts "high fall risk," the meta-model uses these inputs and calculates the final fall risk result based on the learned model and training data. This ensemble-based approach outperforms traditional rule-based models because it can continuously learn from previous predictions and adapt to new data, thus improving its flexibility and reliability. A significant advantage of this technology is its high accuracy, enabling its integration into real-time health monitoring systems. The meta-model integrates fuzzy logic (AI1) and deep belief networks (AI2), enabling intelligent and comprehensive assessment of fall risk based on physiological markers (vital signs) and behavioral markers (impairment in daily activities). The proposed model can be implemented in wearable medical technologies and home automation systems to track and send immediate fall risk alerts to caregivers and healthcare professionals in real time. The system's interactive architecture allows each AI model to enhance the meta-model's decision-making capabilities, forming a highly adaptive and specific predictive mechanism that learns from new data. The meta-model's ability to compare and coordinate AI predictions makes it a powerful tool for predicting fall risk in older adults, thereby promoting early intervention and significantly reducing the incidence of fall-related injuries and associated healthcare costs. AI-driven collaborative models are key to developing predictive healthcare solutions and provide a scalable and intelligent approach to predicting fall risk in older adults. This model aims to predict future falls and assess different levels of fall risk for adults aged 60 and older. Fall risk is categorized into low, medium, and high levels based on a fall risk prediction analysis using the AI-based collaborative meta-model, which informs the prediction results. Data on older adults is collected from public repositories for training and testing purposes.

[0177] Artificial intelligence-based collaborative meta-model fall risk prediction algorithm:

[0178] The AI-based collaborative meta-model fall risk prediction algorithm is constructed by combining the results generated by AI1 and AI2 models. Each individual model is weighted, and then a final composite risk value is calculated. Training labels are set based on the composite risk variable and the maximum functional state (MFS). Here, the meta-model is trained, and the final fall risk prediction is based on learning from stored data (past data) and current input (present data).

[0179] Artificial intelligence-based collaborative meta-model fall risk prediction algorithm:

[0180] Input mapping: Let XAI1 = AI1; XAI2 = AI2

[0181] Assign weights WAI1 and WAI2 to the output:

[0182]

[0183]

[0184] in, :normal; :Low; :medium; :high; :urgent

[0185] Comprehensive Risk Score: The comprehensive risk score Scombined is calculated as follows: Scombined = (WAI1 + WAI2) / 2

[0186] Risk level definition: Risk level R is defined based on Scombined:

[0187]

[0188] Training a random forest meta-model involves a dataset containing training features: Xtrain = {WAI1, WAI2} • training labels; Ytrain = MFS-based risk levels { , , }

[0189] Random Forest Model: Initialize T, the number of decision trees. n_estimators=100; max_depth=10; min_samples_split=4.

[0190] For each tree t in T:

[0191] Randomly select a subset Xt from Xtrain;

[0192] Construct a decision tree using Xt and Ytrain;

[0193] The data is split at each node using the feature {WAI1, WAI2};

[0194] The predictions from all T trees are aggregated through majority voting;

[0195] Predicting fall risk: For test data Xtest={WAI1,WAI2};

[0196] Each test instance is passed through all T decision trees;

[0197] The predicted risk level Rpred is calculated based on the majority vote of T trees.

[0198] For each participant i, the output is: participant i: {Scombined, Rpred};

[0199] A fall risk prediction model based on fuzzy logic principles is specifically designed to identify older adults at risk of falls. The primary motivation for developing this model is to create an effective and useful fall prediction system that helps reduce the economic burden and personal hardship associated with fall-related injuries in older adults. By continuously monitoring vital signs, the model is linked to physiological parameters to detect early biases that may indicate a higher fall risk. Unlike existing methods that rely on multiple health parameters, this model relies solely on vital signs, making it a small yet highly efficient fall risk prediction model. These figures demonstrate the model's efficiency and consistency in classifying fall risk levels with minimal false positives or false negatives. Various studies on fall risk assessment emphasize that most falls in older adults can be prevented through early intervention.

[0200] A recommended fall risk prediction model based on deep belief networks (DBNs), specifically designed for the older market, aims to monitor fall risk by tracking behavioral patterns in certain daily activities (ADLs) in real time. This model assesses fall risk solely based on behavioral data and is able to accurately classify high-risk and non-high-risk patients.

[0201] A collaborative meta-model for AI-based fall risk assessment represents a significant breakthrough in preventative elderly care. Unlike traditional fall monitoring systems that only react after a fall has occurred, this model predicts fall risk by examining vital signs and activities of daily living (ADLs) in real time, making it a predictive model. By integrating fuzzy logic (AI1) for physiological assessment and a deep belief network (AI2) for behavioral assessment within a meta-learning paradigm, the model demonstrates superior predictive efficiency over time, accurately identifying individuals truly at risk without false positives or false negatives.

[0202] The AI1 model in this invention uses a fuzzy rule-based fall prediction algorithm (FPA) to predict falls in the elderly. It monitors vital signs and categorizes fall risk into five levels: , , , and .

[0203] Blood pressure: The first sign of a health problem is often abnormal blood pressure, whether high or low.

[0204] Heart rate: Changes in blood pressure can affect heart rate, causing it to rise or fall outside the normal range.

[0205] Blood oxygen (O2): Fluctuations in heart rate may cause blood oxygen levels in older adults to drop below the normal range of 95-98%.

[0206] Changes in blood pressure can affect heart rate in older adults. Based on background analysis and expert advice, a thorough examination of vital sign patterns was conducted, and 111 fuzzy rules were developed to determine the risk level of future falls in older adults. The risk levels are categorized as follows:

[0207] This indicates there is no risk of falling.

[0208] This indicates whether a fall is possible or unlikely.

[0209] This indicates that if this situation continues, there is a risk of falling in the future.

[0210] This indicates a high risk of falling; seek medical help if necessary.

[0211] This indicates a very high probability of a fall (a fall could occur within seconds or minutes) and will send an alert to caregivers, family members, and emergency services.

[0212] The fall prediction algorithm (Algorithm A1) is based on a set of vital signs data.

[0213] Algorithm A1. Fall Prediction Algorithm (FPA):

[0214] Input: blood pressure, heart rate, and blood oxygen saturation.

[0215] Output: Fall prediction results - risk level classification

[0216] Apply the dataset to the extraction of the input dataset.

[0217] The retrieved features are input into the fuzzy prediction model for training, and rules for classifying risk levels are formulated.

[0218] (a) Blurring: Converting clear data into fuzzy data.

[0219] (b) Membership function: based on parameter values ​​( , , , and Define the membership function. Input function: 'trapmf'.

[0220] (c) Fuzzy rule generation: Formulate fuzzy rules from fuzzy data.

[0221] Fuzzy rule: Rule: For normal blood pressure,

[0222] If pulse pressure is normal

[0223] And heart rate is normal

[0224] And blood oxygen saturation is normal.

[0225] The risk level is normal (no risk of falling was detected).

[0226] Rule: For high blood pressure (126-130 / 80-90).

[0227] If pulse pressure is high

[0228] Heart rate is high (110-129 beats per minute)

[0229] Blood oxygen saturation is moderate (95-96%).

[0230] The risk level is high (if this situation continues, there is a risk of falling within a few days).

[0231] (d) Defuzzification: Converting fuzzy rules into clear rules.

[0232] Membership functions (MFs) range from 0 to 1, with each fall risk level assigned a specific weight: normal -0.2, low -0.4, moderate -0.6, high -0.8, and urgent -1.0. For example, if blood pressure is 90 / 60, the pulse pressure (PP) given by the systolic (S_v) and diastolic (D_v) values ​​is 30. Based on this, a sample prediction rule is as follows:

[0233] If blood pressure is 90 / 60, pulse pressure is less than 40, heart rate is 102, and blood oxygen saturation is 95, then the risk of falling is 0.6.

[0234] If blood pressure is 120 / 80, pulse pressure is 40, heart rate is 90, and blood oxygen saturation is 97, then the risk of falling is 0.2.

[0235] Fuzzy membership functions transform these numerical "if" conditions into linguistic variables (e.g., low, moderate, high). For example,

[0236] If blood pressure is high, pulse pressure is low, heart rate is moderate, and blood oxygen saturation is moderate, then the risk of falling is moderate.

[0237] All numerical decision rules are first processed by converting them into fuzzy rules using membership functions, and then evaluated using fuzzy process analysis. Finally, defuzzification is applied to generate fall risk categories.

[0238] The activity capability conversion algorithm is as follows:

[0239] Input: Raw data from 3D accelerometers

[0240] Outputs: Activity level (sitting, standing, walking, running, jumping) and unknowns

[0241] initialization:

[0242] Clock frequency = 100 scale marks per second

[0243] Window size = 100

[0244] Data Frame → Data Frame

[0245] Time Analysis:

[0246] For each entry in the "Time (Elapsed)" column

[0247] Analysis time (hours:minutes:seconds) = hours × 3600 + minutes × 60 + seconds

[0248] Remove rows containing NaN.

[0249] Feature extraction:

[0250] For each sliding window of size n in the data frame:

[0251]

[0252]

[0253]

[0254]

[0255]

[0256]

[0257] Activity categories: sitting, standing, walking, running, jumping

[0258] The proposed Deep Belief Network (DBN) model consists of three Restricted Boltzmann Machine (RBM) layers: RBM_1, RBM_2, and RBM_3, with 256, 200, and 150 hidden units, respectively. These values ​​are optimized through iterative trials to improve accuracy. Each RBM is trained using contrastive divergence, employing a bottom-up, layer-by-layer (greedy) approach. After training, the output of each RBM is used as the input of the next.

[0259] Input weights are based on fall risk-related activities:

[0260] Sitting: 0.1

[0261] Standing: 0.2

[0262] Walking: 0.3

[0263] Running: 0.6

[0264] Jump: 0.8

[0265] History of setbacks (Yes = 1, No = 0)

[0266] Risk categories are calculated using the following weights

[0267] Low risk: 0.3

[0268] Medium risk: 0.6

[0269] High risk: 0.9

[0270] Each RBM layer captures high-order features through unsupervised learning. After pre-training, the learned weights are used to initialize a deep neural network, forming a stacked deep belief network (DBN) (Algorithm A3). During supervised fine-tuning, early layers may be frozen to preserve learned features, and the model is trained on labeled fall risk data using the Adam optimizer. Only the unfrozen layers are updated, and a softmax classifier produces the final risk output, scaled to low (0.3), medium (0.6), or high (0.9).

[0271] DBN-based fall prediction algorithm (DBN-FPA):

[0272] Input: Daily Activities (ADLs)

[0273] Output: Risk level classification - fall prediction results

[0274] 1. Extract data from input parameters

[0275] 2. Feature extraction → Input the extracted data into the DBN-based prediction model.

[0276] (a) Data standardization

[0277] (X) → Input data with features (fall history, ADLs).

[0278] Perform min-max normalization on each feature in (X):

[0279] X′=(Xmax−Xmin) / (X−Xmin)

[0280] Where Xmin and Xmax are the minimum and maximum values ​​of X, respectively.

[0281] (b) Encoded label

[0282] Use label encoding to transform Y → target label

[0283] (c) Standardize the data → X′

[0284] Encoding labels → Y data is divided into training set and test set.

[0285] (d) Pre-training using a restricted Boltzmann machine

[0286] Training RBM_1:

[0287] Number of components = 256, learning rate = 0.01, number of iterations = 10

[0288] Training RBM_2:

[0289] Number of components = 200, learning rate = 0.01, number of iterations = 10

[0290] Training RBM_3:

[0291] Number of components = 150, learning rate = 0.01, number of iterations = 10

[0292] (e) Fine-tuning → Neural Network + Adam Optimizer

[0293] (f) Initialize the weights of the neural network using the weights of the pre-trained RBMs:

[0294] Model_layers[0] → Set Weights → RBM_1

[0295] Model_layers[1] → Set Weights → RBM_2

[0296] Model_layers[2] → Set Weights → RBM_3

[0297] 3. Training using deep belief networks – risk calculation

[0298] 4. Perform cross-validation using the test dataset → Determine the model's accuracy, specificity, and sensitivity.

[0299] 5. Results evaluation → Validation of findings.

[0300] Backpropagation enables deep networks to learn complex patterns and plays a crucial role in building deep learning models, namely convolutional neural networks and recurrent neural networks.

[0301] RBM_1 (First Hidden Layer - Low-level Features)

[0302] Input: Raw data (sitting, standing, walking, running, and jumping).

[0303] Output: Extract low-level features from the input.

[0304] RBM_2 (Second Hidden Layer - Intermediate Feature)

[0305] Input: Features learned from RBM_1.

[0306] Output: Recognizes higher-order relationships between features.

[0307] RBM_3 (Third Hidden Layer - Advanced Features)

[0308] Input: Features learned by RBM_2.

[0309] Output: Complex patterns were found in the dataset.

[0310] In this invention, all users have obtained informed consent, and strict data privacy and security measures have been implemented to protect sensitive information.

[0311] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, the phrase "comprising an element defined as..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0312] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their likenesses.

Claims

1. A smart fall monitor, characterized in that, include: Wearable data acquisition device, data processing and transmission module, control unit; The control unit includes: A rule-based threshold-based rapid triggering layer monitors events when acceleration exceeds a preset upper threshold or falls below a preset lower threshold, and simultaneously, a body posture angle exceeds a safe range. When these conditions are met, an event flagged as a suspected fall triggers a background AI verification layer. This layer extracts sensor data from a preset time window before and after the event and inputs it into a pre-trained deep neural network classifier to distinguish between a real fall and similar daily activities. If the intelligent verification layer classifies the event as a fall, the system immediately activates an emergency response mechanism. The AI ​​verification layer includes: The AI1 model is a fall risk prediction model based on fuzzy logic, which analyzes users' vital signs data and outputs a first risk level or score. The AI2 model is a fall risk prediction model based on deep belief networks. It is used to analyze users' daily life activity data and output a second risk level or score. The meta-model is configured to receive the outputs of AI1 and AI2 models and construct an ensemble learning classifier using the random forest algorithm. This classifier automatically learns the non-linear relationships and intrinsic patterns between the outputs of the two sub-models to determine the final, comprehensive fall risk level.

2. The intelligent fall detection device as described in claim 1, characterized in that, Wearable data collection devices are used to continuously collect users' vital signs and daily activity data, including: Physiological sensor used to periodically measure blood pressure, pulse pressure, heart rate, and blood oxygen saturation; The motion sensor, including at least one three-axis accelerometer and gyroscope, is used to acquire three-dimensional acceleration and attitude angle data of the body in real time; The data processing and transmission module is located in the wearable data acquisition device and is used to filter and preprocess the acquired raw data and wirelessly transmit the data to the computing platform. A collaborative intelligent prediction and analysis platform, deployed in the cloud or on a local server, contains at least two collaborative artificial intelligence models and a meta-model. The early warning response and user interaction module, based on the comprehensive fall risk prediction results or real-time fall monitoring results, sends corresponding level of early warning or alarm information to the user and / or their caregiver via a mobile application.

3. The intelligent fall detection device as described in claim 2, characterized in that, The formula for determining whether the acceleration exceeds a preset upper limit threshold is: Fall Detection = { True, if a x > UFT or a x < LFT; False, otherwise} a x For the acceleration x-axis component: represents the x-axis acceleration value collected in real time from the inertial measurement unit. During the fall process, the human body may experience weightlessness or severe impact, resulting in a x Mutation occurs, and this variable is used to capture the abnormal characteristics of linear motion; UFT is the upper threshold: a preset upper limit of acceleration, when a x > UFT, it indicates that a high-speed movement or a strong impact may have occurred; LFT is a lower threshold: a preset lower limit value of acceleration, when a x < LFT, which usually corresponds to a weightless state; If the acceleration condition is met, the algorithm will not immediately issue an alarm, but will instead initiate a second verification process. At this point, the gyroscope data is used to analyze the user's body posture angle. After a fall, a person's body posture will change from upright or sitting to an abnormal posture. The algorithm will check whether the user's posture angle exceeds the safe range. The system only determines a fall event when both conditions are met: acceleration exceeding a threshold and abnormal attitude angle.

4. The intelligent fall monitor of claim 3, wherein, Real-time data acquisition and preprocessing: The sensor continuously acquires data, which is then filtered and denoised. The filtering algorithm includes: In the filtering algorithm, the number of buffers N is set, X is the new sampled value; A is the set maximum sampling deviation value, the size of which needs to be judged based on experience; h is set as the minimum number of consecutive samples in the buffer whose deviation is less than A; After filling the buffer, and sorting it; each time a new sample value X is monitored, the new sample value X is subtracted from the historical sample values in the buffer, wherein For the i (i = 1, 2…N) sample value in the buffer, the sample maximum deviation is calculated as: ; find corresponding to denoted , the new sample value X is replaced , so that the data that is not conducive to the output criterion is replaced, and the case that the effective signal is determined as an interference signal and discarded when the signal jumps is prevented. The data in the buffer is sorted again, taking out (i=1,2....N-1) and difference (j=i+1,2...N), if the difference value meets and , record , where is the number of continuous sampling to achieve the deviation less than A, where , sampling value in the buffer; extracted the maximum value of (k=1,2….N) , if <h, then discard; if >h, then output the average value of the corresponding sampling sequence , ​ ； wherein , is a sample number, the filtered result of this time as the final output value of the sampling.

5. The intelligent fall monitor of claim 4, wherein, The fall risk prediction method based on fuzzy logic implemented by the AI1 model includes the following steps: Input steps: Obtain the user's vital signs data, including blood pressure, pulse pressure, heart rate, and blood oxygen saturation; Fuzzification steps: Convert clear numerical inputs into linguistic variables using trapezoidal membership functions; Reasoning steps: The fuzzy data is input into a fuzzy rule base containing at least 111 preset rules for evaluation. The rules are used to classify the user's fall risk into five levels: normal, low, medium, high, and emergency. Defuzzification step: Convert the fuzzy inference result into a clear risk score or rating label output; Output steps: The predicted fall risk score or level is used as the output of the AI1 model for further decision-making by the meta-model.

6. A smart fall detector as described in claim 5, characterized in that, The fall risk prediction method based on deep belief networks implemented by the AI2 model includes the following steps: Input processing steps: Based on three-dimensional accelerometer data, identify the user's daily activities, including sitting, standing, walking, running, and jumping, and assign a preset weight related to the importance of each activity. Combine this with fall history records to form a feature vector. Model building steps: Build a deep belief network containing multiple stacked restricted Boltzmann machine layers, perform layer-by-layer greedy pre-training, and use the Adam optimizer for supervised fine-tuning on labeled data; Risk classification steps: The network learns the intrinsic connection between user activity patterns and fall risk, and finally classifies users' fall risk into three levels: low, medium and high, and outputs the corresponding level or score; Output steps: The predicted fall risk level or score is output as the AI2 model for further decision-making by the meta-model.

7. A smart fall detector as described in claim 6, characterized in that, The meta-model-based collaborative fall risk prediction method includes the following steps: Input mapping step: Receive the output of AI1 model (XAI1) and the output of AI2 model (XAI2), and optionally assign dynamic or static weights (WAI1 and WAI2) to each input. Meta-classifier construction and training steps: Using the outputs of AI1 and AI2 models as input features, and Morse Fall Score (MFS) risk levels (low, medium, high) as labels, train a random forest classifier; Prediction and decision-making steps: For new input features, prediction is performed using a trained random forest meta-model, and the final, unified fall risk level (Rpred) is determined by majority voting of all decision trees. Output steps: Output a combined risk score (Scombined) and / or a final fall risk prediction rating (Rpred) for each user.

8. A smart fall detector as described in claim 7, characterized in that, Artificial intelligence-based collaborative meta-model fall risk prediction algorithm: Input mapping: Let: X AI1=AI1; XAI2=AI2 Assign weights WAI1 and WAI2 to the output: ； in, :normal; :Low; :medium; :high; :urgent; Comprehensive Risk Score: The comprehensive risk score Scombined is calculated as follows: Scombined = (WAI1 + WAI2) / 2; Risk level definition: Risk level R is defined based on Scombined:

9. A smart fall detector as described in claim 8, characterized in that, Training a random forest meta-model includes training features Xtrain={WAI1,WAI2} • training labels; Ytrain = risk level based on MFS. Random Forest Model: Initialize T, number of decision trees, n_estimators=100 max_depth=10 min_samples_split=4; For each tree t in T: Randomly select a subset Xt from Xtrain; Build a decision tree using Xt and Ytrain; Split the data at each node using the feature {WAI1, WAI2}; The predictions from all T trees are aggregated through majority voting; Predicting fall risk: For test data Xtest={WAI1,WAI2}; Each test instance is passed through all T decision trees; The predicted risk level Rpred is calculated based on the majority vote of T trees; For each participant i, the output is: participant i: {Scombined, Rpred}.

10. The intelligent fall detection device as described in claim 9, characterized in that: ‌ Long-term, non-invasive fall risk assessment is conducted by continuously monitoring vital signs and activity patterns. When a high risk of fall is predicted, an alert is issued to the user and their care network, prompting them to take preventative measures. When a real-time fall event is detected, an emergency alarm is automatically triggered, sending an alert containing the user's location information.

Images