Search

CN-122025186-A - Multi-dimensional evaluation system for physical activity change track of colorectal cancer chemotherapy patient

CN122025186ACN 122025186 ACN122025186 ACN 122025186ACN-122025186-A

Abstract

The invention relates to the technical field of analysis of physical activity change tracks, in particular to a multi-dimensional data acquisition and feature extraction module which acquires multi-dimensional data, extracts static feature vectors and time sequence feature vectors, a track pattern mining and anchor point identification module which adopts a time sequence analysis method to extract overall fluctuation features and local fluctuation features in the time sequence feature vectors, fuses the overall fluctuation features and the local fluctuation features to obtain track feature vectors, divides a plurality of track feature vectors into different clusters, each cluster is a type of physical activity change track to obtain physical activity change tracks in different patterns, and physical activity change anchor points, a personalized model training and physical activity prediction module which uses physical activity change track patterns as layering variables, classifies the static feature vectors and the time sequence feature vectors of corresponding patients according to track patterns to generate a plurality of track pattern homogeneous groups, and trains a personalized evaluation model based on the plurality of track pattern homogeneous groups.

Inventors

  • WANG YUWEI

Assignees

  • 中国人民解放军总医院第八医学中心

Dates

Publication Date
20260512
Application Date
20260228

Claims (10)

  1. 1. A multi-dimensional assessment system for a trace of physical activity of a colorectal cancer chemotherapeutic patient, comprising: The multi-dimensional data acquisition and feature extraction module (100) acquires physiological data and clinical diagnosis and treatment data of each chemotherapy patient in the whole chemotherapy period to form multi-dimensional data, and extracts static feature vectors and time sequence feature vectors of the patients in the whole chemotherapy period in the multi-dimensional data; The track pattern mining and anchor point identification module (200) adopts a time sequence analysis method to extract integral fluctuation characteristics and local fluctuation characteristics in time sequence characteristic vectors, specifically adopts a quadratic polynomial fitting method to capture the integral fluctuation characteristics, adopts a sliding window analysis method to identify the local fluctuation characteristics and respectively identifies physical energy change anchor points; the method comprises the steps of establishing a mapping relation between local fluctuation characteristics and integral fluctuation characteristics, and fusing the integral fluctuation characteristics and the local fluctuation characteristics to obtain track characteristic vectors, dividing a plurality of track characteristic vectors into different clusters, wherein each cluster is a type of physical energy change track, so that physical energy change tracks in different modes and physical energy change anchor points are obtained; The personalized model training and physical ability prediction module (300) takes the physical ability change track mode as a layering variable, classifies static feature vectors and time sequence feature vectors of corresponding patients according to the track mode to generate a plurality of track mode homogeneous groups, and trains a personalized evaluation model based on the plurality of track mode homogeneous groups, wherein the personalized evaluation model is used for predicting physical ability change track modes of chemotherapeutic patients and physical ability change anchor points in the physical ability change track, and the physical ability evaluation value of the physical ability of each chemotherapeutic stage in the chemotherapeutic process.
  2. 2. The multi-dimensional evaluation system of the physical activity change track of the colorectal cancer chemotherapeutic patient according to claim 1, wherein the track pattern mining and anchor point identification module (200) is characterized by comprising the following steps of: Taking X days before chemotherapy as a unified time step, arranging time sequence feature vectors according to time sequence to form a time sequence feature matrix corresponding to each chemotherapy patient; sequentially extracting integral fluctuation features representing macroscopic indexes and local fluctuation features representing microscopic indexes in each time sequence feature matrix by adopting a combined extraction method, establishing a mapping relation between the local fluctuation features and the integral fluctuation features, and deeply fusing the integral fluctuation features and the local fluctuation features to obtain track feature vectors corresponding to each chemotherapeutic patient; vector similarity among the corresponding track feature vectors of each patient is calculated in a crossing mode, the track feature vectors are classified based on the vector similarity, and physical ability change tracks in different modes are obtained.
  3. 3. The multi-dimensional evaluation system of the physical activity change track of colorectal cancer chemotherapeutic patients according to claim 2, wherein the combined extraction method is specifically as follows: capturing the integral fluctuation characteristics in the time sequence characteristic matrix by adopting quadratic polynomial fitting, namely, constructing a quadratic polynomial fitting model to generate a fitting curve; The method comprises the steps of adopting a sliding window analysis method to identify local fluctuation characteristics, namely, carrying out continuous non-overlapping window division on a time sequence characteristic matrix according to time sequence to obtain a plurality of continuous local time sequence data subsets, calculating local slope of each local data subset, judging local fluctuation modes corresponding to the local data subsets, extracting key time points corresponding to the local fluctuation modes, defining the key time points as physical energy change anchor points, substituting the physical energy change anchor points into a quadratic polynomial fitting model, and outputting fitting values corresponding to the physical energy change anchor points by the quadratic polynomial fitting model to obtain the local fluctuation characteristics; and combining the integral fluctuation feature and the local fluctuation feature according to a fixed sequence to obtain a track feature vector.
  4. 4. The multi-dimensional assessment system of the trace of the physical activity of a colorectal cancer chemotherapeutic patient according to claim 3, wherein the trace feature vector is classified as: Initializing cluster centers, namely setting the number K of the initial cluster centers, randomly selecting K track feature vectors as the initial cluster centers, and calculating the vector similarity between each track feature vector and all the initial cluster centers; Distributing the vector, namely distributing the track feature vector to the center of the initial cluster with the maximum vector similarity to form K initial clusters; the cluster center is updated, namely, the average value vector of all the track feature vectors in each initial cluster is calculated, and the average value vector is used as the new cluster center of the corresponding initial cluster; iterative optimization, namely repeatedly distributing vectors and updating cluster centers until the cluster centers are not changed any more, and converging the clustering result at the moment; and outputting the physical ability change track, wherein after convergence, each cluster is a physical ability change track.
  5. 5. The multi-dimensional evaluation system for physical activity change tracks of colorectal cancer chemotherapeutic patients according to claim 4, wherein the personalized model training and physical activity prediction module (300) receives static feature vectors, time sequence feature vectors, track mode mining and physical activity change tracks of each patient in the multi-dimensional data acquisition and feature extraction module (100), classifies the static feature vectors and time sequence feature vectors of the corresponding patients according to track modes by taking the physical activity change track modes as layering variables, generates a plurality of track mode homogeneous groups, and divides the static feature vectors, the time sequence feature vectors and the physical activity change tracks in each homogeneous group according to fixed proportion to obtain a training set, a verification set and a test set for training the personalized evaluation model.
  6. 6. The multi-dimensional evaluation system of the physical activity change track of the colorectal cancer chemotherapeutic patient according to claim 5, wherein the individual difference modeling branch is used for learning association rules between static feature vectors in a homogeneous group and different physical activity change tracks and different physical activity change anchor points in the homogeneous group, specifically, the static feature vectors in a training set are taken as input, nonlinear association of the static feature vectors and the physical activity change track modes and the physical activity change anchor points in the track mode mining and anchor point identification module (200) is captured, corresponding parameters of the individual difference modeling branch are trained, and after training is completed, a preliminary prediction mode of the physical activity change track and a physical activity change anchor point candidate interval are output.
  7. 7. The multi-dimensional evaluation system of the physical activity change track of the colorectal cancer chemotherapeutic patient according to claim 6, wherein the individual difference modeling branch comprises an input layer, a hidden layer and an output layer; The input layer receives static feature vectors; The hidden layer comprises a plurality of full-connection layers and is used for learning nonlinear association of static feature vectors and physical ability change tracks of different modes in a homogeneous group, physical ability change anchor points in the physical ability change tracks corresponding to each patient, and finally outputting prediction scores corresponding to the physical ability change tracks of the modes and physical ability change anchor point candidate intervals; each full-connection layer comprises a plurality of neurons, each neuron in the full-connection layer is directly connected with all neurons in the previous layer and is used for integrating all information of the previous full-connection layer, and the input of the previous layer is subjected to linear transformation through a weight matrix and a bias term to adjust the influence weights of different characteristics; introducing a nonlinear activation function to the full-connection layer, so that the full-connection layer can learn nonlinear association between static feature vectors and physical energy change tracks of different modes and physical energy change anchor point candidate intervals, and finally outputting corresponding output vectors; The output layer is used for respectively converting the output vector of the hidden layer into a corresponding primary prediction mode and a physical energy change anchor point candidate interval.
  8. 8. The multi-dimensional evaluation system of the physical activity change track of the colorectal cancer chemotherapeutic patient according to claim 7, wherein when the individual difference modeling branches learn the association rule between the static feature vector and the track pattern mining and anchor point identification module (200) in the multi-dimensional data acquisition and feature extraction module (100), the output layer receives the output vector of the last hidden layer, converts the output vector of the hidden layer into the prediction probability of each physical activity change track pattern, and outputs the track pattern with the maximum prediction probability as the preliminary prediction pattern; The method comprises the steps of outputting a physical change anchor point candidate interval when an association rule between a static feature vector and a track pattern in an individual difference modeling branch learning multidimensional data acquisition and feature extraction module (100) and a physical change anchor point in an anchor point identification module (200), wherein an output layer comprises two neurons and each neuron corresponds to a group of independent weight vectors and bias items, each neuron is respectively associated with a hidden layer output, a candidate start day and a candidate stop day, receiving an output vector of a last hidden layer, and outputting a candidate start day and a candidate stop day in the physical change anchor point candidate interval through linear transformation.
  9. 9. The multi-dimensional evaluation system of the physical activity change track of the colorectal cancer chemotherapeutic patient according to claim 8, wherein the time sequence dependent node branches take time sequence feature vectors in a training set as input, four gate control units of LSTM are adopted, hidden states corresponding to each moment in the long-range associated output time sequence feature vectors of the time sequence feature vectors are mined, the hidden states generated at each moment are sequentially arranged according to time sequence, and finally a hidden state sequence is formed; By static feature vectors For inquiring vectors, importance scoring is carried out on the hidden state at each moment to obtain an individual difference weight coefficient set which finally covers the whole chemotherapy cycle; and the individual difference weight coefficient sets are aggregated according to the chemotherapy stage, and the contribution degree of the corresponding time sequence fluctuation of different moments and the chemotherapy stage to the time sequence characteristic of the physical ability evaluation value is analyzed.
  10. 10. The multi-dimensional evaluation system of the physical activity change track of the colorectal cancer chemotherapeutic patient according to claim 9 is characterized in that the personalized evaluation model splices a preliminary prediction result and a hidden state sequence of a track mode to obtain a fusion feature vector; The individual difference weight coefficient set is adopted as dynamic constraint, the physical ability change anchor point candidate interval output by the individual difference modeling branch is accurately contracted, and a specific time point with the highest matching degree with the anchor point mode is screened out and used as a predicted physical ability change anchor point; The method comprises the steps of receiving time sequence feature vectors before chemotherapy of all patients in each physical energy change track, taking an average value of the time sequence feature vectors before chemotherapy of all patients in the same physical energy change track, taking a calculation result as a physical energy baseline value corresponding to each patient in each chemotherapy stage, weighting and fusing an individual difference weight coefficient set and a time sequence feature contribution degree, combining the physical energy baseline value to obtain a single-moment basic evaluation value, quantifying the single-moment basic evaluation value in each stage, and finally obtaining the physical energy evaluation value in each chemotherapy stage.

Description

Multi-dimensional evaluation system for physical activity change track of colorectal cancer chemotherapy patient Technical Field The invention relates to the technical field of analysis of physical activity change tracks, in particular to a multidimensional evaluation system of physical activity change tracks of colorectal cancer chemotherapy patients. Background Colorectal cancer is malignant tumor occurring on the epithelium of colon or rectum mucous membrane, is one of common malignant tumors of digestive system, the incidence part covers colon (including ascending colon, transverse colon, descending colon and sigmoid colon) and rectum, the colorectal cancer is treated by adopting chemotherapy in the prior art, and in the study of CHALLEENGE random control test, the study of physical activity and prognosis association queue of colorectal cancer patients, the study of physical energy change longitudinal tracking in chemotherapy period and the like, the physical activity level and disease recurrence risk, total survival time and life quality improvement of the colorectal cancer patients are obviously positively correlated after the colorectal cancer patients finish chemotherapy, the relative risk of disease recurrence can be reduced by 28% by maintaining moderate physical activity, and the fatigue related to chemotherapy can be effectively relieved, and the heart and lung functions and muscle strength can be improved; Although the above-mentioned researches have clearly elucidated the association between the therapeutic effect after colorectal cancer chemotherapy and the physical activity change track of the patient, the existing clinic has also tried to collect objective physiological data (such as daily steps, energy consumption, heart rate and exercise intensity) related to the physical activity of the patient by a medical-grade intelligent sensor (capturing exercise and physiological signals in real time by means of acceleration and heart rate sensing module), and analyze the physical activity change rule of the patient by combining with multidimensional information such as clinical diagnosis and treatment data, but there is still a core bottleneck in practical application, specifically: The physical performance basis, the toxic and side effect influence degree of the chemotherapeutic drugs and the physical tolerance threshold of patients in different treatment stages (such as chemotherapy administration period, intermittent period and recovery period) are obviously different, if the physical activity reduction amplitude (such as 30% or 50% of the decrease of the number of steps of the administration period compared with the baseline), the recovery potential (such as whether the intermittent period can recover to the physical activity level above the baseline by more than 80%) and the dynamic influence degree of the chemotherapy on the physical performance of the patients (such as the regulation and control effect of the drug metabolism rate on the physical performance recovery) corresponding to each stage cannot be accurately determined; meanwhile, the personalized rehabilitation intervention scheme is formulated by taking the staged physical performance as a core and according to different stages of intervention key points, namely, the administration period is required to be aimed at maintaining basic physical strength to avoid excessive activities and increasing physical burden, the intermittent period is required to be cored by gradually increasing activity strength to assist physical performance recovery, and the recovery period is required to be focused on consolidating activity ability to reduce recurrence risk; If the accurate cognition of the key information is lacking, the intervention intensity, frequency and mode of the current physical energy state adaptation of the patient cannot be judged, so that the follow-up cannot further formulate personalized rehabilitation intervention schemes attached to different treatment stages for the patient, and specifically, the formulated intervention measures either increase the physical burden of the patient due to the overhigh intensity or cannot achieve the effect of improving prognosis due to the insufficient intensity; On the other hand, a key intervention window for physical activity decline (such as an inflection point for the first occurrence of obvious decline of physical activity and a period for suddenly accelerating the decline rate) is the best time for blocking continuous deterioration of physical ability, the accurate identification of the key intervention window is required to rely on real-time control of physical activity decline amplitude and physical ability dynamic influence by chemotherapy, if the key intervention window is not timely identified, the best time for preventive intervention is missed, at the moment, the physical activity level of a patient is lower than a benefit threshold value for a long time (namely, the moderate act