Risk prediction method for patients with Parkinson's dysphagia

Abstract

The invention discloses a risk prediction method for patients with Parkinson's dysphagia, and relates to the technical field of medical health. The method comprises the steps that S1, a multi-modal data fusion framework is built, the multi-modal data fusion framework aims at effectively fusing biomarkers, dynamic swallowing kinematics characteristics and clinical data, the core of the multi-modal data fusion framework comprises characteristic alignment, inter-modal interaction modeling and fusion decision, and a first risk prediction probability is output; through a multi-modal data fusion algorithm, a dynamic swallowing function quantification technology and a lightweight edge deployment scheme, high-precision prediction, early warning and personalized intervention of the Parkinson's dysphagia risk are realized, the method has remarkable clinical value, technical barriers and market competitiveness, and the core demand of medical productization landing is met.

Description

Technical Field

[0001] The present invention relates to the field of medical health technology, and in particular to a risk prediction method for patients with Parkinson's dysphagia. Background Art

[0002] In existing technologies, the assessment of Parkinson's dysphagia mainly relies on clinical scales (such as FOIS, SWAL-QOL) or imaging examinations (such as VFSS, FEES), but they have the following defects:

[0003] Subjectivity: The scale relies on the subjective judgment of patients or doctors and lacks accuracy;

[0004] Lag: Imaging examinations are only used when symptoms are obvious and cannot provide early warning;

[0005] Single dimension: Lack of comprehensive analysis of biomarkers and dynamic swallowing function.

[0006] Therefore, a new solution to the above problems needs to be proposed. Summary of the Invention

[0007] The purpose of the present invention is to provide a risk prediction method for patients with Parkinson's dysphagia to solve the technical problems raised in the background technology.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the risk of Parkinson's dysphagia patients, comprising at least the following steps:

[0009] S1: Build a multimodal data fusion framework. The multimodal data fusion framework aims to effectively integrate biomarkers, dynamic swallowing kinematics characteristics, and clinical data. The core of the multimodal data fusion framework includes feature alignment, inter-modal interaction modeling, and fusion decision-making, and outputs the first risk prediction probability.

[0010] S2: Build a dynamic swallowing function quantification algorithm. Based on the laryngeal sensor signals, extract swallowing coordination, swallowing force, and swallowing pattern abnormalities, and finally generate a second risk prediction probability. The swallowing coordination includes but is not limited to the pharyngeal delay time and the displacement speed of the hyoid-laryngeal complex. The swallowing force is the integral of the EMG signal amplitude. The swallowing pattern abnormality is the frequency of laryngeal micro-movements during the non-swallowing period.

[0011] S3: Set weights based on the first risk prediction probability and the second risk prediction probability to obtain the final risk prediction result.

[0012] Furthermore, the biomarkers include, but are not limited to, detecting α-synuclein oligomers, TNF-α, and IL-6 levels in saliva / blood, combined with miRNA expression profiles;

[0013] The swallowing kinematics is measured by wearable laryngeal sensors, namely EMG + accelerometer, which collect the laryngeal movement trajectory, muscle activation timing and swallowing frequency during swallowing;

[0014] The clinical data include but are not limited to Hoehn-Yahr stage, UPDRS-III score, previous history of pneumonia and drug response data.

[0015] Furthermore, the feature alignment is used to eliminate the dimensional differences and timing offsets of different modal features, and specifically includes the following steps:

[0016] Time series alignment, for dynamic gene motion signals, the sampling frequency f is used s =100Hz, and the dynamic time warping (DTW) algorithm is used to align the sparse time points of clinical data:

[0017]

[0018] Where: X={x1,x2,...,x N}、Y={y1,y2,...,y M} are two sets of modal time series; P is the set of all possible alignment paths; d(x i ,y j )=|x i -y j | 2 is the Euclidean distance metric;

[0019] Feature normalization is used to perform Z-score normalization on biomarkers and clinical scores:

[0020]

[0021] Where: μ k and σ k is the mean and standard deviation of the kth root eigenvalue; z (k) is the normalized feature vector.

[0022] Furthermore, the intermodal interaction modeling is used to capture the nonlinear relationship between biomarkers and kinematic features, and specifically includes the following steps:

[0023] Covariance attention mechanism, defining the inter-modal correlation matrix d1 and d2 are the two-modal feature dimensions:

[0024]

[0025] in, and are the biomarkers and kinematic characteristics of the i-th sample respectively; μbio and μ kinematic are the mean vectors of biomarkers and kinematic features respectively; C mn represents the correlation strength between the m-th biomarker and the n-th kinematic feature;

[0026] Attention weight generation, calculating the inter-modal attention weights based on the correlation matrix:

[0027]

[0028] where is a trainable projection matrix; [·] is the feature concatenation operation; α mm is the interaction weight between biomarker m and kinematic feature n.

[0029] Furthermore, the fusion decision is used to map multi-modal information into a unified risk score, specifically including the following steps:

[0030] Tensor fusion, constructing a high-order tensor to capture the combined effects between modes:

[0031]

[0032] where h bio and h kineumatic are biomarkers and kinematic features respectively; is the tensor outer product operation; h clin represents clinical features; is the joint effect representing biomarker i, kinematic feature j, and clinical feature k;

[0033] Low-rank approximation for dimensionality reduction, to avoid dimensional explosion, Tucker decomposition is used to reduce the dimensionality of the tensor:

[0034]

[0035] where is the core tensor (r i <d i ); are the factor matrices of each mode; × i is the product of the tensor along the i-th mode;

[0036] The first risk probability output, inputting the dimensionality-reduced tensor features into a fully connected layer:

[0037] p = σ(w T .vec(G) + b)

[0038] where vec(G) is to flatten the core tensor into a vector; w T and b are trainable parameters; σ is the Sigmoid function.

[0039] Further, the S2 at least includes the following steps:

[0040] Input data, where the EMG signal is used to extract the swallowing force; the motion sensor signal is used to extract the swallowing coordination; the sound or vibration sensor signal is used to analyze abnormal swallowing patterns;

[0041] Swallowing coordination, which includes the pharyngeal delay time and the displacement velocity of the hyoid-laryngeal complex. By processing the laryngeal sensor signal, these two features are extracted;

[0042] The pharyngeal delay time is determined by detecting the characteristic points of the laryngeal sensor signal to obtain the time of the start of swallowing and comparing it with the time of the end of the pharyngeal phase;

[0043] The displacement velocity of the hyoid-laryngeal complex uses a sensor or accelerometer to monitor the movements of the hyoid bone and the larynx, and calculates the displacement velocity of the hyoid-laryngeal complex through the time series of the motion signal;

[0044] Step: Extract the displacement of the signal and calculate the displacement velocity of the hyoid bone and the larynx during swallowing;

[0045] The swallowing force is measured by the amplitude integration of the EMG signal. The muscle activity during swallowing will generate fluctuations in the EMG signal, and the swallowing force can be quantified by integrating the amplitude of the EMG signal, that is

[0046] Extract the EMG signal; filter the signal, and use high-pass filtering to remove low-frequency noise; calculate the amplitude integration of the EMG signal;

[0047] Abnormal swallowing patterns are identified by analyzing the micro-movement frequency of the larynx during the non-swallowing period. By processing the laryngeal sensor signal, the micro-movement frequency characteristics during the non-swallowing state are extracted, and these frequency changes may indicate abnormal patterns during swallowing;

[0048] The micro-movement frequency of the larynx during the non-swallowing period uses Fourier transform to analyze the spectrum of the non-swallowing period signal and identify the micro-movement frequency therein;

[0049] The generation of the risk prediction probability is based on the above three features, namely swallowing coordination, swallowing force, and abnormal swallowing patterns, and generates the second risk prediction probability through a machine learning model.

[0050] Further, the S3 at least includes the following steps:

[0051] Input data:

[0052] p1 is the risk prediction probability output based on the multi-modal data fusion framework;

[0053] p2 is the risk prediction probability output by the dynamic swallowing function quantification algorithm;

[0054] α is the weight of the first risk prediction, α∈[0,1];

[0055] Calculate the final risk prediction and combine the two risk prediction results according to the weighted average method:

[0056] p final =α·p1+(1-α)·p2

[0057] Further optimize the weights and use cross-validation to adjust the weights to obtain the best fusion effect;

[0058] Output the final risk prediction probability p final .

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] This invention achieves high-precision prediction, early warning and personalized intervention of Parkinson's dysphagia risk through multimodal data fusion algorithm and dynamic swallowing function quantification technology deployment scheme. It has significant clinical value, technical barriers and market competitiveness, and meets the core needs of medical product implementation. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0062] Figure 1 It is a schematic diagram of the overall process of the present invention. DETAILED DESCRIPTION

[0063] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0064] See also Figure 1 A method for predicting the risk of Parkinson's disease in patients with dysphagia comprises at least the following steps:

[0065] S1: Build a multimodal data fusion framework. The multimodal data fusion framework aims to effectively integrate biomarkers, dynamic swallowing kinematics characteristics, and clinical data. The core of the multimodal data fusion framework includes feature alignment, inter-modal interaction modeling, and fusion decision-making, outputting the first risk prediction probability.

[0066] S2: Build a dynamic swallowing function quantification algorithm. Based on laryngeal sensor signals, extract swallowing coordination, swallowing force, and abnormal swallowing patterns, and finally generate a second risk prediction probability. Swallowing coordination includes, but is not limited to, the pharyngeal delay time and the displacement velocity of the hyoid-laryngeal complex. Swallowing force is the integral of the EMG signal amplitude, and abnormal swallowing patterns are the micro-motion frequencies of the larynx during non-swallowing periods.

[0067] S3: Set weights based on the first risk prediction probability and the second risk prediction probability to obtain the final risk prediction result.

[0068] Biomarkers include, but are not limited to, detecting the levels of α-synuclein oligomers, TNF-α, and IL-6 in saliva / blood, and combining miRNA expression profiles;

[0069] Swallowing kinematics collects the laryngeal movement trajectory, muscle activation timing, and swallowing frequency during swallowing through a wearable laryngeal sensor, namely EMG + accelerometer;

[0070] Clinical data includes, but is not limited to, Hoehn-Yahr staging, UPDRS-Ⅲ score, previous history of pneumonia, and drug response data.

[0071] Feature alignment is used to eliminate the dimensional differences and temporal offsets of different modality features, and specifically includes the following steps:

[0072] Time series alignment. For dynamic gene motion signals, the sampling frequency f s = 100Hz is used, and for sparse time points in clinical data, the dynamic time warping (DTW) algorithm is used for alignment:

[0073]

[0074] where: X = {x1, x2,..., x N}, Y = {y1, y2,..., y M} are two groups of modality time series; P is the set of all possible alignment paths; d(x i , y i ) = |x i - y j | 2 is the Euclidean distance metric;

[0075] Feature standardization is used to perform Z-score standardization on biomarkers and clinical scores:

[0076]

[0077] where: μ k and σ k are the mean and standard deviation of the k-th eigenvalue; z(k) is the standardized eigenvector.

[0078] Inter-modal interaction modeling is used to capture the non-linear associations between biomarkers and kinematic features, and specifically includes the following steps:

[0079] Covariance attention mechanism to define the inter-modal correlation matrix d1 and d2 are the feature dimensions of the two modalities:

[0080]

[0081] where and are the biomarker and kinematic feature of the i-th sample respectively; μ bio and μ kinematic are the mean vectors of the biomarker and kinematic feature respectively; C mn represents the correlation strength between the m-th biomarker and the n-th kinematic feature;

[0082] Attention weight generation, calculating the inter-modal attention weights based on the correlation matrix:

[0083]

[0084] where is the trainable projection matrix; [·] is the feature concatenation operation; α mm is the interaction weight between the biomarker m and the kinematic feature n.

[0085] Fusion decision is used to map multi-modal information into a unified risk score, and specifically includes the following steps:

[0086] Tensor fusion, constructing a high-order tensor to capture the combined effect between patterns:

[0087]

[0088] where h bio and h kineumatic are the biomarker and kinematic feature respectively; is the tensor outer product operation; h clin represents the clinical feature; is the combined effect representing the biomarker i, kinematic feature j, and clinical feature k;

[0089] Low-rank approximation for dimensionality reduction, to avoid dimensional explosion, Tucker decomposition is used to reduce the dimensionality of the tensor:

[0090]

[0091] where is the core tensor (ri <d i ); is the factor matrix for each mode; × i is the product of the tensor along the i-th mode;

[0092] The first risk probability output, input the tensor features after dimensionality reduction into the fully connected layer:

[0093] p = σ(w T .vec(G) + b)

[0094] where, vec(G) is to flatten the core tensor into a vector; w T and b are trainable parameters; σ is the Sigmoid function.

[0095] S2 includes at least the following steps:

[0096] Input data, the EMG signal is used to extract the swallowing force; the motion sensor signal is used to extract the swallowing coordination; the sound or vibration sensor signal is used to analyze the abnormal swallowing pattern;

[0097] Swallowing coordination, the swallowing coordination includes the pharyngeal latency and the displacement velocity of the hyoid-laryngeal complex. By processing the laryngeal sensor signal, these two features are extracted;

[0098] The pharyngeal latency determines the start time of swallowing by detecting the characteristic points of the laryngeal sensor signal and compares it with the end time of the pharyngeal phase;

[0099] The displacement velocity of the hyoid-laryngeal complex monitors the movement of the hyoid bone and the larynx using a sensor or an accelerometer, and calculates the displacement velocity of the hyoid-laryngeal complex through the time series of the motion signal;

[0100] Step: Extract the displacement of the signal and calculate the displacement velocity of the hyoid bone and the larynx during swallowing;

[0101] The swallowing force is measured by the amplitude integration of the EMG signal. The muscle activity during swallowing will cause fluctuations in the EMG signal, and the swallowing force can be quantified by integrating the amplitude of the EMG signal, that is

[0102] Extract the EMG signal; filter the signal, and use high-pass filtering to remove low-frequency noise; calculate the amplitude integration of the EMG signal;

[0103] The abnormal swallowing pattern is identified by analyzing the micro-motion frequency of the larynx during the non-swallowing period. By processing the laryngeal sensor signal, the micro-motion frequency characteristics in the non-swallowing state are extracted, and these frequency changes may indicate an abnormal pattern during swallowing;

[0104] During the non-swallowing period, the micro-movement frequency of the larynx is analyzed using Fourier transform to identify the micro-movement frequency in the non-swallowing period signal;

[0105] Generation of the risk prediction probability: Based on the above three features, namely swallowing coordination, swallowing force, and abnormal swallowing pattern, a machine learning model is used to generate a second risk prediction probability.

[0106] S3 includes at least the following steps:

[0107] Input data:

[0108] p1 is the risk prediction probability output based on the multi-modal data fusion framework;

[0109] p2 is the risk prediction probability output by the dynamic swallowing function quantification algorithm;

[0110] α is the weight of the first risk prediction, and α ∈ [0, 1];

[0111] Calculate the final risk prediction: According to the weighted average method, fuse the two risk prediction results:

[0112] p final = α·p1 + (1 - α)·p2

[0113] Further optimize the weight: Use cross-validation to adjust the weight to obtain the best fusion effect;

[0114] Output the final risk prediction probability p final .

[0115] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, in any aspect, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be encompassed within the present invention. Any reference signs in the claims should not be regarded as limiting the claimed rights involved.

Claims

1. A risk prediction method for Parkinson's patients with dysphagia, characterized in that: At least include the following steps: S1: Build a framework for multimodal data fusion, which aims to effectively fuse biomarkers, dynamic swallowing kinematic features, and clinical data. The core of the multimodal data fusion framework includes feature alignment, cross-modal interaction modeling, and fusion decision-making, and outputs the first risk prediction probability; S2: Build a dynamic swallowing function quantification algorithm. Based on laryngeal sensor signals, extract swallowing coordination, swallowing force, and abnormal swallowing patterns, and finally generate the second risk prediction probability. The swallowing coordination includes, but is not limited to, the pharyngeal delay time and the displacement velocity of the hyoid-laryngeal complex. The swallowing force is the integral of the EMG signal amplitude, and the abnormal swallowing pattern is the micro-motion frequency of the larynx during non-swallowing periods; S3: Set weights based on the first risk prediction probability and the second risk prediction probability to obtain the final risk prediction result.

2. The risk prediction method for Parkinson's dysphagia patients according to claim 1, wherein: The biomarkers include, but are not limited to, detecting the levels of α-synuclein oligomers, TNF-α, and IL-6 in saliva / blood, and combining miRNA expression profiles; The swallowing kinematics is collected through a wearable laryngeal sensor, namely EMG + accelerometer, to obtain the laryngeal movement trajectory, muscle activation timing, and swallowing frequency during swallowing; The clinical data includes, but is not limited to, the Hoehn-Yahr stage, UPDRS-Ⅲ score, previous pneumonia history, and drug response data.

3. The risk prediction method for Parkinson's patients with dysphagia according to claim 2, wherein: The feature alignment is used to eliminate the dimensional difference and temporal offset of different modal features, and specifically includes the following steps: Time series alignment, with a sampling frequency f s = 100 Hz for dynamic gene motion signals and used for sparse time points in clinical data. The dynamic time warping algorithm is used for alignment: Where: X = {x1, x2,..., x N}, Y = {y1, y2,..., y M} are two sets of modal time series; P is the set of all possible alignment paths; d(x i , y j ) = |x i - y j | 2 is the Euclidean distance metric; Feature standardization is used to perform Z-score standardization on biomarkers and clinical scores: where: μ k and σ k are the mean and standard deviation of the k-th eigenvalue; z (k) is the eigenvector after standardization.

4. A risk prediction method for Parkinson's patients with swallowing disorders according to claim 3, characterized in that: The cross-modal interaction modeling is used to capture the non-linear association between biomarkers and kinematic features, and specifically includes the following steps: Covariance attention mechanism, defining the inter-modal correlation matrix d1 and d2 are the feature dimensions of the two modalities: wherein, and are the biomarker and kinematic feature of the i-th sample respectively; μ bio and μ kinematic are the mean vectors of the biomarker and kinematic feature respectively; C mn represents the correlation strength between the m-th biomarker and the n-th kinematic feature; Attention weight generation, calculating cross-modal attention weights based on the correlation matrix: Among them, is a trainable projection matrix; [·] is a feature concatenation operation; α mm is the interaction weight between biomarker m and kinematic feature n.

5. The risk prediction method for Parkinson's dysphagia patients according to claim 4, characterized in that: The fusion decision-making is used to map multimodal information into a unified risk score, and specifically includes the following steps: Tensor fusion, constructing a high-order tensor to capture the combined effect between modes: where h bio and h kineumatic are biomarker and kinematic feature respectively; is the tensor outer product operation; h clin represents the clinical feature; is for representing the combined effect of biomarker i, kinematic feature j and clinical feature k; Low-rank approximation for dimensionality reduction. To avoid dimensional explosion, Tucker decomposition is used to reduce the dimensionality of the tensor: Among them, is the core tensor (r i <d i )); is the factor matrix of each mode; × i is the product of the tensor along the i-th mode; Output of the first risk probability, inputting the dimension-reduced tensor features into a fully connected layer: p = σ(w T .vec(G)+b) where \(vec(G)\) is the core tensor flattened into a vector; \(w\) T and \(b\) are trainable parameters; \(\sigma\) is the Sigmoid function.

6. A risk prediction method for Parkinson's patients with dysphagia according to claim 1, characterized in that: The S2 at least includes the following steps: Input data, the EMG signal is used to extract swallowing force; the motion sensor signal is used to extract swallowing coordination; the sound or vibration sensor signal is used to analyze abnormal swallowing patterns; Swallowing coordination, which includes the pharyngeal delay time and the displacement velocity of the hyoid-laryngeal complex. These two features are extracted by processing laryngeal sensor signals; The pharyngeal delay time is determined by detecting the feature points of the laryngeal sensor signal to determine the start time of swallowing and comparing it with the end time of the pharyngeal phase; The displacement velocity of the hyoid-laryngeal complex monitors the movement of the hyoid bone and larynx using a sensor or accelerometer, and calculates the displacement velocity of the hyoid-laryngeal complex through the time series of the motion signal; Step: Extract the displacement of the signal and calculate the displacement velocity of the hyoid bone and larynx during swallowing; The swallowing force is measured by the amplitude integration of the EMG signal. The muscle activity during swallowing generates fluctuations in the EMG signal, and the swallowing force can be quantified by integrating the amplitude of the EMG signal, which is Extract the EMG signal; filter the signal, and use high-pass filtering to remove low-frequency noise; calculate the amplitude integration of the EMG signal; Abnormal swallowing patterns are identified by analyzing the micro-motion frequency of the larynx during the non-swallowing period. By processing the larynx sensor signal, the micro-motion frequency characteristics during the non-swallowing state are extracted, and these frequency changes may indicate abnormal patterns during swallowing; For the micro-motion frequency of the larynx during the non-swallowing period, use Fourier transform to analyze the spectrum of the non-swallowing period signal and identify the micro-motion frequency therein; The generation of the risk prediction probability is based on the above three characteristics, namely swallowing coordination, swallowing force, and abnormal swallowing patterns, and a machine learning model is used to generate the second risk prediction probability.

7. A risk prediction method for Parkinson's patients with dysphagia according to claim 1, characterized in that: The S3 at least includes the following steps: Input data: p1 is the risk prediction probability output based on the multi-modal data fusion framework; p2 is the risk prediction probability output by the dynamic swallowing function quantification algorithm; α is the weight of the first risk prediction, and α ∈ [0, 1]; Calculate the final risk prediction, and fuse the two risk prediction results according to the weighted average method: p final = α·p1 + (1 - α)·p2 Further optimize the weight, and use cross-validation to adjust the weight to obtain the best fusion effect; Output the final risk prediction probability p final .

Images