Prediction system for early Parkinson's disease

Abstract

The invention relates to a prediction system for an early Parkinson's disease. The prediction system is characterized by comprising a task issuing module, a data acquisition module and a data analysis module, wherein the task issuing module issues a test task instruction to a user at the front end of the equipment; the data acquisition module is used for acquiring measurement parameters of specified behaviors through acquisition equipment according to different test tasks; and the data analysis module is used for carrying out statistics and analysis on measurement parameters based on the prediction model and carrying out early Parkinson's disease prediction. The optimal screening tool for the early Parkinson's disease is deduced on the basis of cooperation of multi-modal acquisition equipment and a prediction system for the early Parkinson's disease, early screening of the Parkinson's disease is achieved, and the clinical problem that early diagnosis of the Parkinson's disease is difficult is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent medical treatment, in particular to a prediction system for early Parkinson's disease. BACKGROUND

[0002] Parkinson's disease (PD) early motor signs show specificity, stability and objectivity, so accurate measurement of early PD motor signs, combined with artificial intelligence technology, screening of sensitive and stable digital biological indicators, and establishment of an effective early PD screening model will provide a solution for early diagnosis of PD.

[0003] However, in the process of studying early diagnosis of PD, it is found that early PD motor signs show diversity, that is, motor slowing can occur in different parts of the body, specifically in gait characteristics (reduced amplitude of upper arm and forewall swing, swing, walking drag, step speed, turning time, etc.), hand and foot fine motor characteristics (asymmetry in finger and foot movement speed and amplitude, sequence effect), voice characteristics (low-pitched voice, slow speech, prolonged pause), facial expression changes (decreased blinking frequency, decreased facial expression muscle movement amplitude), and eye scanning changes (decreased scanning and anti-scanning speed and amplitude, increased anti-scanning error rate, etc.). That is, a PD patient can have multiple motor signs. In addition, early PD motor signs also have heterogeneity, that is, the affected parts and severity differ among individuals, resulting in different motor signs in different patients.

[0004] The existing reported PD prediction models have the following defects: ① Most prediction models are constructed from single-mode measured motor parameters, which come from the measurement of a certain motor function under a single action task. The prediction model based on single mode has poor generalization ability and is difficult to solve the problem of heterogeneity of early PD motor signs. ② Recently, some studies have used facial expression features and voice features in free monologue tasks, or combined facial expression and pointing action in fixed paragraph reading with back-and-forth straight walking tasks to establish dual-mode and triple-mode PD prediction models. The increase in modalities improves the ability of the model to identify different motor function impairments in PD, however, the modalities in these models are randomly selected, rather than data-driven, and the performance of predicting early PD has not reached an excellent level, with an AUC value of only 0.84 (dual-mode) and 0.82 (triple-mode).

[0005] In the aspect of motion feature measurement technology, the existing detection technology mainly completes the parameter measurement of gait, facial expression, finger tapping, sound and eye scanning through wearing motion sensing units, monocular vision system, plantar pressure sensing, recording pen, eye scanning instrument and other ways. The monocular vision system based on three-dimensional human posture estimation (using mobile phones and tablets) can complete the measurement of gait and facial expression features, but this technology relies on data-driven method to promote two-dimensional posture to three-dimensional, and the posture estimation accuracy is greatly limited by training data, and there is an inherent defect of depth visual ambiguity; the plantar pressure sensing device can only measure the gait features of both lower limbs, and cannot measure the motion features of upper limbs, trunk and facial expression. The existing researches either use wearable measurement devices to complete gait feature measurement, or use monocular vision system to complete gait and / or facial expression feature measurement, and there are also researches using multiple collection devices to complete the measurement of gait, facial expression, sound, scanning and other motion features, resulting in too many measurement instruments, complex process and long time consumption, poor patient compliance, and current researches are limited to researches. In clinical practice, these technologies cannot meet the simple and efficient demand of predicting early PD. SUMMARY

[0006] The present application aims to overcome the above-mentioned defects, and provides a prediction system for early Parkinson's disease, which is characterized by:

[0007] The prediction system for early Parkinson's disease provided by the present application comprises a task issuing module, a data acquisition module and a data analysis module.

[0008] The task issuing module issues test task instructions to the users in front of the equipment.

[0009] The data acquisition module acquires measurement parameters of specified behaviors through acquisition equipment according to different test tasks.

[0010] The data analysis module statistically analyzes the measurement parameters based on an early Parkinson's disease prediction model to predict early Parkinson's disease.

[0011] Further, the prediction system for early Parkinson's disease provided by the present application is characterized by:

[0012] The construction method of the early Parkinson's disease prediction model is:

[0013] S1. Set the motion tasks to be tested, and acquire the motion parameters of early Parkinson's disease patients and healthy control population under each motion task.

[0014] S2. Compare all motion parameters between early Parkinson's disease patients and healthy control population, and screen out the motion features belonging to early Parkinson's disease, and construct a mode with the motion features belonging to early Parkinson's disease.

[0015] S3. In combination with the modalities of S2, a machine learning method is used to build a full-modal early Parkinson's disease prediction model.

[0016] Further, the application provides an early Parkinson's disease prediction system, which is characterized in that:

[0017] The optimal modal screening is performed on the full-modal early Parkinson's disease prediction model, and the modal that performs best in terms of single-modal prediction ability, stability and contribution value of full-modal prediction is screened out.

[0018] Further, the application provides an early Parkinson's disease prediction system, which is characterized in that:

[0019] The early Parkinson's disease prediction model is built based on the screened single modal.

[0020] Further, the application provides an early Parkinson's disease prediction system, which is characterized in that:

[0021] The test task instruction includes eye saccade and anti-saccade, fixed passage reading and 4-meter straight-line walking back and forth. Further, the application provides an early Parkinson's disease prediction system, which is characterized in that:

[0022] The measurement parameters of eye saccade and anti-saccade include saccade and anti-saccade reaction time, saccade and anti-saccade amplitude and speed, saccade and anti-saccade correct times, saccade and anti-saccade error times, saccade and anti-saccade non-response times, saccade and anti-saccade correct rate, and saccade and anti-saccade error rate.

[0023] Further, the application provides an early Parkinson's disease prediction system, which is characterized in that:

[0024] The measurement parameters of fixed passage reading include blink rate, mean and standard deviation of eyebrow lifting, mean, standard deviation and left-right asymmetry index of eyebrow tilting, mean, standard deviation and left-right asymmetry index of eye and eyebrow shape, palpebral aperture, mean and standard deviation of upper and lower lip movement, mean and standard deviation of lip height, mean and standard deviation of lip foot adduction and abduction, and asymmetry index.

[0025] Further, the application provides an early Parkinson's disease prediction system, which is characterized in that:

[0026] 4m back-and-forth straight line walking measurement parameters include shoulder inclination mean and standard deviation, wrist swing amplitude mean and standard deviation, wrist swing amplitude coefficient of variation and left-right asymmetry, wrist swing speed mean and standard deviation, wrist swing speed coefficient of variation and left-right asymmetry, forearm swing angle mean and standard deviation and left-right asymmetry index, calf swing angle mean and standard deviation and left-right asymmetry index, foot lifting distance mean and standard deviation, foot lifting distance coefficient of variation and left-right asymmetry, gait cycle, step speed, step frequency, step width and turning time.

[0027] Further, the early Parkinson's disease prediction system provided by the present application is characterized in that:

[0028] The acquisition device comprises a binocular vision acquisition system.

[0029] Further, the early Parkinson's disease prediction system provided by the present application is characterized in that:

[0030] The acquisition device is based on a human body 3D pose estimation algorithm of RSB-POSE, and can capture and measure large-scale and small-scale motion characteristics of the human body, and position large and small joints of the limbs to accurately measure the motion direction and speed of the large and small joints and changes thereof.

[0031] The acquisition device is based on PFLD technology to obtain 98 facial marker points and pupil marker points from the recorded video, and complete acquisition of dynamic facial expression information and eye saccade motion parameters.

[0032] Effects and advantages of the present application:

[0033] The present application provides the best screening tool for early Parkinson's disease patients based on the multi-modal acquisition device in cooperation with the early Parkinson's disease prediction system, realizes early Parkinson's disease screening, and solves the clinical problem of early Parkinson's disease diagnosis difficulty. Specifically, the present system compares a plurality of motion tasks commonly used at present, proposes an optimal modal selection strategy, determines that gait, facial expression parameters in fixed paragraph reading and saccades are the best modal in early Parkinson's disease prediction, which is superior to other 7 commonly used modalities, and gives a summary conclusion in the current Parkinson's disease diagnosis artificial intelligence research field. DETAILED DESCRIPTION

[0034] The present application can be implemented in various modifications and can have various embodiments. However, this is not intended to limit the present application to a specific embodiment, but should be understood to include all modifications, equivalents and even alternatives falling within the spirit and technical scope of the present application.

[0035] Test example:

[0036] I. Research object:

[0037] This study included 151 patients with early-stage PD who met the clinical diagnostic criteria for PD, had a disease duration of <5 years, and a HY grade ≤2; and 138 healthy controls were also included.

[0038] The study subjects were divided into a modeling cohort (n=239, PD group: 126; healthy (HC) group: 113) and a validation cohort (n=50, PD group: 25; HC group: 25) in a 5:1 ratio.

[0039] Sample size calculation: The modeling sample size was calculated using the confidence interval method based on the area under the ROC curve. The AUC value reached 0.85, 1-Alpha = 0.95, the number of cases and controls was equal, and the confidence interval width should not exceed 0.15. Considering a possible 5% sample testing loss, the minimum sample size per group in the test cohort was calculated using PASS software (21.0) to be 56 cases. The ratio of the total sample size in the validation cohort to the total sample size in the test cohort should be at least 1:5.

[0040] II. Research Steps and Results

[0041] Based on clinical experience, this embodiment includes 11 movement tasks and 186 motor characteristic parameters, requiring participants to complete them as instructed. Patients with Parkinson's disease (PD) are required to complete the tasks at least 12 hours after discontinuing their PD medication. All studies were conducted in a single examination room. Multimodal intelligent data acquisition equipment was used to measure the motor characteristics. Before completion, the display automatically showed the task requirements, demonstration movements, and voice explanation. Researchers could provide guidance as needed to ensure participants understood and correctly performed the movements.

[0042] This multimodal intelligent acquisition device is a binocular vision acquisition system. Based on the RSB-POSE human 3D pose estimation algorithm, it captures and measures large-scale and small-scale motion features of the human body, and accurately measures the direction, speed, and changes of motion of the large and small joints of the limbs. Based on PFLD technology, it acquires 98 facial and pupil marker points from the recorded video to complete the acquisition of dynamic facial expression information and eye saccade motion parameters.

[0043] 1. Exercise tasks and parameters

[0044] (1) Spontaneous blinking: In a quiet, noise-free environment, the subject sat with his head fixed in a U-shaped support. A multimodal intelligent robot was placed 0.55m directly in front of him, and a 19-inch monitor was placed 0.70m away at a 45° angle in front of him. Under natural light, the monitor displayed a white background with a solid gray dot in the center. The patient was asked to focus on the central dot for 2.5 minutes.

[0045] Measurement parameters: blink rate, blink cycle (single) and total duration (right eye), eyelid closure phase time (left and right eyes), eye opening phase time (left and right eyes), average amplitude-speed ratio of eye closing phase (left and right eyes), and amplitude-speed ratio of eye opening phase (left and right eyes).

[0046] (2) Sagging and Anti-sagging: Under the same testing environment and position, the patient fixates on the signal cursor on the screen with both eyes. After 9-point calibration, the patient completes saccade and anti-sagging fixation as required. Each group consists of 30 repetitions, and the last 20 repetitions are used for data analysis and parameter measurement.

[0047] The measurement parameters include: saccade and antisaccade reaction time, saccade and antisaccade amplitude and speed, number of correct saccades and antisaccades, number of incorrect saccades and antisaccades, number of no response saccades and antisaccades, saccade and antisaccade accuracy rate, and saccade and antisaccade error rate.

[0048] (3) Facial expressions during the fixed-paragraph reading task: In a quiet, noise-free environment, the subject sits and looks straight ahead, reading aloud the children's song "Under the Bridge in Front of the Door" displayed on the screen. The patient can try reading it once, and after confirming that they can complete the reading task, they can start recording once.

[0049] Data collected included: blink rate, mean and standard deviation of eyebrow raising (right eye), mean, standard deviation, and left-right asymmetry index of eyebrow tilt (left and right), mean, standard deviation, and left-right asymmetry index of eyebrow shape (left and right), palpebral aperture (left and right), mean and standard deviation of upper and lower lip movements, mean and standard deviation of lip height, and mean and standard deviation of lip foot adduction and abduction, and asymmetry index (left and right).

[0050] (4) Facial expressions during the free monologue task: In a quiet, noise-free environment, the subject sits and looks straight ahead. They can choose one of three topics from the prompts on the screen to complete the free monologue task, with a sampling time of 1 minute (childhood anecdotes, hobbies, family).

[0051] The collected parameters are the same as those for facial expressions in a fixed paragraph reading task.

[0052] (5) Continuous vowel: In a quiet, noise-free environment, the subject sits and looks straight ahead. According to the task requirements displayed on the screen, the subject relaxes and, as if speaking normally, takes a deep breath and continuously pronounces: Ah, (first tone) (repeated three times, pausing slightly after each repetition); after resting, takes a deep breath and continuously pronounces: Eh, (demonstration, similar to the English word 'a') (repeated three times, pausing slightly after each repetition).

[0053] Parameter acquisition: F0, Jitter, Shimmer, CPP median CPP p25 CPP p75, RFA, HNR.

[0054] (6) Alternating pronunciation: In a quiet, noise-free environment, the subject sits and looks straight ahead. According to the task requirements displayed on the screen, the subject relaxes and, as if speaking normally, takes a deep breath and repeats the following in one breath: pa / ta / ka (repeat as many times as possible in one breath, with the instructor counting. After reaching 12 repetitions, the subject can be signaled to stop. The task is completed 3 times in total).

[0055] Parameter acquisition: DDK rate and stability, VOT, RFA.

[0056] (7) Fixed paragraph reading: In a quiet, noise-free environment, the subject sits and looks straight ahead. According to the task requirements displayed on the screen, the subject relaxes and reads the children's song "Under the Bridge in Front of the Door" once, as if speaking normally.

[0057] Parameter acquisition: mean and variability of sound intensity, NSR, DPI median DPI p25 DPI p75 、F0、RFA.

[0058] (8) Free Monologue: In a quiet, noise-free environment, the subject sits and looks straight ahead. Following the task requirements displayed on the screen, the subject relaxes and, as if speaking normally, chooses one of three topics from the prompts on the screen to complete a free monologue task. The sample is taken for 1 minute (childhood anecdotes, hobbies, family).

[0059] Parameter acquisition: sound duration, mean and standard deviation of DPI, mean and variability of sound intensity, NSR, DPI median DPI p25 DPI p75 、F0、RFA.

[0060] (9) Finger tapping: In a quiet, seated position, the subject is asked to tap continuously with their thumb and forefinger, repeating 10 times. Requirements: The fingers should be spread as wide as possible, and the tapping speed should be as fast as possible (refer to section 4 of MDS-UPDRS-III for instructions).

[0061] Parameters collected: tapping speed (left and right), speed variation trend (left and right) and left-right asymmetry index, amplitude left-right asymmetry index, amplitude variation trend (left and right) and left-right asymmetry index.

[0062] (10) Toe-tapping exercise: In a quiet, seated position, the subject's heels should remain on the ground throughout. The subject should raise their toes and tap the ground, repeating 10 times. Requirements: Raise the toes as high as possible and tap the ground as quickly as possible (refer to exercise description 7 in MDS-UPDRS-III).

[0063] Parameters collected: tapping speed (left and right), speed variation trend (left and right) and left-right asymmetry index, amplitude left-right asymmetry index, amplitude variation trend (left and right) and left-right asymmetry index.

[0064] ⑪ 4-meter round trip straight walking: In a quiet state, the patient stands at the starting point (1.5 meters away from the robot), walks straight ahead, walks naturally for 4 meters, then goes around the marker (turns left) and returns to the starting point.

[0065] Parameters collected: mean and standard deviation of shoulder tilt (left and right), mean and standard deviation of wrist swing amplitude (left and right), coefficient of variation of wrist swing amplitude (left and right) and left-right asymmetry, mean and standard deviation of wrist swing speed (left and right), coefficient of variation of wrist swing speed (left and right) and left-right asymmetry, mean and standard deviation of forearm swing angle (left and right) and left-right asymmetry index, mean and standard deviation of lower leg swing angle (left and right) and left-right asymmetry index, mean and standard deviation of foot lift distance (left and right), coefficient of variation of foot lift distance (left and right) and left-right asymmetry, gait cycle, gait speed, cadence, stride width, and turning time.

[0066] 2. PD motion feature screening and modality construction

[0067] (1) PD Feature Screening: In the test cohort, all motor parameters were compared between the PD and HC groups. A total of 78 features showed differences, including 8 saccadic features, 2 spontaneous blinking features, 14 facial expression features for fixed paragraph reading tasks, 18 facial expression features for free monologue tasks, 17 gait features, 3 finger tapping features, 1 toe tapping feature, 3 voice features of continuous vowels, 3 voice features of alternating pronunciation, 3 voice features of fixed paragraph reading, 6 voice features of free monologue, and 17 gait features.

[0068] (2) PD modality construction: There are 11 modalities in total, and each modality contains different motion features under the motion task.

[0069] 3. Construction and validation of a full-modal early PD prediction model

[0070] In the test queue, four machine learning methods (logistic regression, support vector machine, random forest, and multilayer perceptron) were used in combination with 11 modalities to construct four full-modal early PD prediction models. The AUC values ​​of each model in the test set were calculated using 10-fold cross-validation. The full-modal early PD prediction model constructed using MLP performed best, with the highest AUC value of 0.89 and an accuracy of 0.82. In the validation queue, the AUC value of the full-modal early PD prediction model constructed using MLP was calculated to be 0.88, with an accuracy of 0.76.

[0071] 4. Optimal mode selection, construction and validation of early PD prediction models with simplified modes

[0072] (1) Optimal Mode Selection: Our team has set up an optimal mode selection strategy. A mode is considered optimal if it meets the following three conditions:

[0073] In the test queue,

[0074] A.SHAP analysis shows that this mode makes a significant contribution to decision-making in the full-modal model;

[0075] B. Single-modality has the ability to distinguish between PD and HC in all machine learning analysis methods;

[0076] C. Single-modal analysis has the ability to reliably distinguish between PD and HC among all machine learning analysis methods.

[0077] Based on this strategy, facial expressions and gait in saccades and fixed paragraph reading were ultimately selected as the best modalities because they met the above three conditions.

[0078] (2) Establishment and validation of a streamlined and efficient early PD prediction model

[0079] Twenty-eight streamlined early PD prediction models were constructed using four machine learning methods, with three optimal modalities randomly combined in a unimodal, bimodal, and trimodal manner. The AUC values ​​of each model in the test set were calculated using 10-fold cross-validation. The results showed that among these 28 models, the early prediction model combining MLP and trimodality had the highest AUC value (0.87) and an accuracy of 0.78, and was thus identified as the best early PD prediction model. This model had an AUC value of 0.84 and an accuracy of 0.76 in the validation queue.

[0080] 5. Comparison between simplified and efficient early PD prediction models and clinician judgment

[0081] Three senior Parkinson's disease (PD) specialists (over 15 years of experience) from our center who did not participate in the study, three resident physicians from our center (around 3 years of experience), and three general practitioners from three community hospitals in Yangpu District (over 15 years of experience) made judgments on PD and HC by watching fixed-segment reading videos of facial expressions and gait (each subject's facial expression and gait videos were combined) of all subjects in the de-identified validation set. The results showed that the accuracy of the 3-modal MLP prediction model (0.76) was similar to that of the senior PD specialists (0.77), and both were higher than the accuracy of the resident physicians (0.72) and community physicians (0.68).

[0082] Based on the above research findings, we propose that the three tasks ultimately identified in this study are simple in action, do not rely on specialized knowledge, are time-consuming (within 3 minutes), are entirely contactless, and have no special environmental requirements. The intelligent robot can provide efficient real-time diagnostic suggestions through specific parameters in these three tasks. Its simplicity and efficiency make it suitable as a screening tool for early PD screening in communities, offering a solution to the difficulties in early PD diagnosis.

[0083] The specific parameters that need to be extracted in these three tasks are as follows:

[0084] 1. Gait parameters

[0085] shoulder tilt angle

[0086] The specific method for obtaining it is as follows:

[0087] S1. Obtain the coordinates of the left and right shoulders (the marker points are the joint points detected by the human pose estimation algorithm, which can be understood as the center points of the movable joints at the shoulders);

[0088] S2. Connect the left and right coordinates;

[0089] S3. Calculate the angle between the line and the horizontal plane, and take the absolute value.

[0090] wrist swing range

[0091] The specific method for obtaining it is as follows:

[0092] S1. Obtain the coordinates of the wrist marker point in each frame;

[0093] S2. Calculate the vertical difference between wrist markers during a gait cycle (excluding body height, which is standardized).

[0094] wrist swing speed

[0095] The specific method for obtaining it is as follows:

[0096] S1. Obtain the time when the wrist marker reaches the highest and lowest points for each arm swing;

[0097] S2. Calculate the time difference for each arm swing in S1;

[0098] S3. Extract the vertical difference of wrist marker points / the time difference between reaching the highest and lowest points in a gait cycle;

[0099] Forearm swing angle

[0100] The specific method for obtaining it is as follows:

[0101] S1. Obtain the maximum flexion and extension angles of the forearm relative to the torso in the sagittal plane during each arm swing, where positive is when the arm swings in front of the body and negative is when the arm swings behind the body;

[0102] S2. Extract the maximum flexion and extension angles of the forearm relative to the trunk in the sagittal plane during a gait cycle, and calculate the difference between them.

[0103] Lower leg swing angle

[0104] The specific method for obtaining it is as follows:

[0105] S1. Obtain the maximum flexion and extension angles of the thigh relative to the torso in each step;

[0106] S2. Extract the maximum flexion and extension angles of the thigh relative to the trunk in the sagittal plane during a gait cycle, and calculate the difference between them.

[0107] Foot elevation distance

[0108] The specific method for obtaining it is as follows:

[0109] S1. Obtain the marker point of the heel or toe in each step;

[0110] S2. Calculate the vertical height of the foot off the ground in each step;

[0111] S2. Calculate the average, maximum, and minimum height of heel lift-off from the ground during one gait cycle (excluding body height, which is standardized).

[0112] Turning time:

[0113] The specific method for obtaining it is as follows:

[0114] S1. Extract the rotated image;

[0115] S2. When the turning angle changes by more than 10° relative to the initial angle, the turning is considered to have started. When the turning angle changes by more than 170° relative to the initial angle, the turning is considered to have been completed. The joint points of the left and right shoulders (defined as the same as the shoulder tilt) are used as markers to determine the turning. The time taken for the turning process is recorded as the 10°-170° turning time (Sec).

[0116] 2. Sagging and Anti-sagging Parameters

[0117] Salivation reaction time

[0118] The specific method for obtaining it is as follows:

[0119] Calculate the time interval from the appearance of the target point to the start of eye saccade (with the pupil as the marker point), in milliseconds.

[0120] Anti-sagittal reaction time

[0121] The specific method for obtaining it is as follows:

[0122] Calculate the time interval from the appearance of the target point to the start of the eye's reverse saccade movement, expressed in seconds.

[0123] Sagging and counter-sagging amplitude

[0124] The specific method for obtaining it is as follows:

[0125] Calculate the maximum pupillary movement between the start and end positions of the eye during a single saccade and antisaccade, in percentages (%).

[0126] Sagging and counter-sagging amplitude

[0127] The specific method for obtaining it is as follows:

[0128] Calculate the maximum speed of pupil movement between the start and end of a single saccade and antisaccade, expressed as % / s.

[0129] Number of responses (saccades / counter-saccades)

[0130] The specific method for obtaining it is as follows:

[0131] S1. Count the number of times the saccade (anti-saccade) direction is correct;

[0132] S2. Count the number of incorrect saccade (anti-saccade) directions;

[0133] S2. Count the number of times a saccade (anti-saccade) did not elicit a response;

[0134] S3. Count the number of times the saccade (anti-saccade) direction was correct / the total number of saccades (anti-saccades).

[0135] S4. Count the number of times the saccadic (anti-saccadic) direction was reversed / the total number of saccadic (anti-saccadic) times.

[0136] 3. Facial expressions

[0137] Asymmetry: The calculation formula is: (left eye parameter - right eye parameter) / left eye parameter

[0138] The function and effect of this embodiment:

[0139] Regarding the data acquisition equipment, a binocular vision acquisition system (resolution: 1280×960, 30fps frame rate) is configured. It utilizes an RSB-POSE-based 3D human pose estimation algorithm to capture and measure large and small-scale human motion features, comprehensively covering the large and small joints of the limbs, and accurately measuring the direction, speed, and changes in joint movement. PFLD technology is used to acquire 98 facial and pupil markers from recorded video, completing the acquisition of dynamic facial expression information and eye saccade motion parameters. The equipment is also equipped with a sound transmission system and an image display system, which can automatically guide patients to complete instructed actions, simultaneously acquiring patient video and audio information. This equipment can perform gait, facial expression, sound, eye saccades, and fine motor movements of the hands and toes in a contactless manner. Voice and video instructions are provided during the detection process to improve automation and simplify the measurement process. The binocular vision acquisition system improves the accuracy and efficiency of parameter measurement, enhances the subject's experience, and provides technical support for realizing a simple and efficient early PD screening tool.

[0140] Regarding the prediction system, considering that bradykinesia is a necessary condition for the clinical diagnosis of PD, this study conducted a case-control study, including early PD patients (disease duration <5 years) and HC (Hospital Hypersensitivity to Disease) patients, establishing a test cohort and a validation cohort. Based on the experience of Parkinson's disease experts, 11 motor tasks were designed to measure the motor parameters of the study subjects (PD patients who had not received treatment or had stopped medication for more than 12 hours) when performing the corresponding actions, totaling 186 motor features, covering eye movements, facial expressions, voice, gait, and fine motor skills of fingers and toes. This is the first comprehensive measurement of bradykinesia-related features in multiple parts of the human body. In the test set, we identified 78 early PD-related motor features through inter-group comparisons, constructing 11 modalities. We used four machine learning methods to build four early PD prediction models based on the full modality (11 modalities). The AUC values ​​of all models exceeded 0.85 on both the test and validation sets, with the full modality prediction model combining multilayer perceptron (MLP) achieving the highest AUC (test set: 0.89; validation set: 0.88). Building upon this foundation, our team innovatively proposed an optimal modality screening strategy. Gait, facial expression, and saccades in the fixed-paragraph monologue task demonstrated the best performance in terms of single-modality diagnostic capability, stability, and contribution to overall modality diagnosis, outperforming the other eight modalities. The early PD prediction model constructed using MLP combined with the three modalities achieved the highest AUC value (test set: 0.87; validation set: 0.84), containing 38 motion parameters. Based on these analysis results, we ultimately proposed a simplified early PD prediction model, consisting of three actions (4-meter straight-line back-and-forth walking, 1-minute fixed-paragraph reading, and 1-minute eye saccades and anti-saccades task), with a detection time of 3 minutes. This model analyzes 38 motion parameters across the three tasks to determine whether a subject is an early PD patient, achieving an accuracy of 78% (test set) and 74% (validation set), respectively. This is superior to the judgments of community doctors and neurology residents (with approximately 3 years of experience) and close to the judgments of experts with over 15 years of clinical experience in Parkinson's disease treatment. We believe that the early PD prediction model paired with this multimodal intelligent robot is simple and efficient, and can be used for early PD screening to solve the clinical challenge of early PD diagnosis.

[0141] While the foregoing has focused on embodiments, these are merely illustrative and do not limit the invention. Those skilled in the art will understand that various modifications and applications not illustrated above can be made without departing from the essential characteristics of these embodiments. For example, the constituent elements specifically shown in the embodiments can be implemented through modifications. Furthermore, various differences related to such modifications and applications should be interpreted as being included within the scope of the invention as defined in the appended claims.

Claims

1. A predictive system for early Parkinson's disease, characterized in that: It includes a task publishing module, a data acquisition module, and a data analysis module; The task issuing module issues test task instructions to users at the front end of the device. The data acquisition module collects measurement parameters of specified behaviors through acquisition devices according to different test tasks; The data analysis module, based on an early prediction model for Parkinson's disease, statistically analyzes and measures parameters to predict early-stage Parkinson's disease.

2. The early prediction system for Parkinson's disease as described in claim 1, characterized in that: The method for constructing the early prediction model for Parkinson's disease is as follows: S1. Set up the exercise tasks to be tested and collect the exercise parameters of Parkinson's disease patients and healthy control groups under each exercise task; S2. Compare all motor parameters between the Parkinson's disease group and the healthy control group to screen out motor characteristics belonging to Parkinson's disease, and construct a mode based on the motor characteristics belonging to Parkinson's disease. S3. By combining machine learning methods with the modalities of S2, a full-modal early prediction model for Parkinson's disease was constructed.

3. The early prediction system for Parkinson's disease as described in claim 2, characterized in that: Optimal modality screening was conducted for early prediction models of Parkinson's disease across all modalities, identifying the modality that performed best in terms of single-modal prediction ability, stability, and contribution to all-modal prediction.

4. The early prediction system for Parkinson's disease as described in claim 3, characterized in that: An early prediction model for Parkinson's disease was constructed based on the selected single modalities.

5. The early prediction system for Parkinson's disease as described in claim 1, characterized in that: The test task instructions include eye saccades and anti-saccades, fixed paragraph reading, and 4-meter round-trip straight-line walking.

6. The early prediction system for Parkinson's disease as described in claim 5, characterized in that: The measurement parameters for eye saccades and antisaccades include saccade and antisaccade reaction time, maximum saccade and antisaccade amplitude and speed, number of correct saccades and antisaccades, number of correct saccades and antisaccades, number of no-response saccades and antisaccades, saccade and antisaccade error rate, and saccade and antisaccade accuracy rate.

7. The early prediction system for Parkinson's disease as described in claim 5, characterized in that: The measurement parameters for the fixed paragraph reading include blink rate, mean and standard deviation of eyebrow raising, mean, standard deviation and left-right asymmetry index of eyebrow tilt, mean, standard deviation and left-right asymmetry index of eyebrow shape, eyelid aperture, mean and standard deviation of upper and lower lip movement, mean and standard deviation of lip height, mean and standard deviation of lip foot adduction and abduction, and asymmetry index.

8. The early prediction system for Parkinson's disease as described in claim 5, characterized in that: The measurement parameters for the 4-meter round-trip straight walking include the mean and standard deviation of shoulder inclination, the mean and standard deviation of wrist swing amplitude, the coefficient of variation and left-right asymmetry of wrist swing amplitude, the mean and standard deviation of wrist swing speed, the coefficient of variation and left-right asymmetry of wrist swing speed, the mean and standard deviation of forearm swing angle and the left-right asymmetry index, the mean and standard deviation of lower leg swing angle and the left-right asymmetry index, the mean and standard deviation of foot lift distance, the coefficient of variation and left-right asymmetry of foot lift distance, gait cycle, gait speed, cadence, stride width, and turning time.

9. The early prediction system for Parkinson's disease as described in claim 1, characterized in that: The acquisition device includes a binocular vision acquisition system.

10. The early prediction system for Parkinson's disease as described in claim 1, characterized in that: The acquisition device, based on the RSB-POSE human 3D pose estimation algorithm, completes the capture and measurement of large-scale and small-scale motion features of the human body, and accurately measures the direction and speed of the movement of the large and small joints of the limbs and their changes. The acquisition device, based on PFLD technology, obtains 98 facial markers and pupil markers from the recorded video, completing the acquisition of dynamic facial expression information and eye saccade motion parameters.