Neurotoxicity grade evaluation method and system based on font analysis
Through multimodal data collection and AI models based on font analysis, the neurotoxicity level of patients' handwriting is automatically assessed, solving the problems of low assessment accuracy and high medical pressure in existing technologies, and achieving efficient and accurate ICANS diagnosis and early warning.
Patent Information
- Application Number
- CN202510499325.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-21
- Publication Date
- 2025-09-19
AI Technical Summary
In existing technologies, patients' handwritten records lack objectivity, resulting in low accuracy of ICANS grade assessment and high intensity of medical work, especially during CAR-T cell therapy, which requires intensive dynamic monitoring.
A font analysis-based method is used to collect multimodal writing data through the "pen tip" sensor, and the AI model is combined to automatically output the ICE score, including data collection, automatic feature engineering, model training and application, to build a neurotoxicity level assessment system.
It achieves efficient and objective ICANS grade assessment, reduces medical workload, provides diagnostic trend analysis and early warning, and improves assessment accuracy and efficiency.
Smart Images
Figure CN120674043A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of multimodal neurotoxicity assessment, and in particular to a data toxicity level assessment method and system based on font analysis. Background Art
[0002] Cellular immunotherapy monitoring is crucial in the medical field. In particular, the multimodal sensing-based automated neurotoxicity assessment system plays a key role in the diagnosis and prevention of ICANS in the CAR-T treatment ward of a tertiary hospital's hematology department. This system provides healthcare workers with an accurate and efficient diagnostic tool, reducing workload and improving patient outcomes.
[0003] Existing technologies for evaluating patients' handwritten records have two major technical drawbacks:
[0004] 1. Low Accuracy: Currently, records of patient handwriting only provide result data, while accurate assessment of a patient's grade requires the handwriting process itself. Existing data cannot objectively reflect the strength, duration, or presence of handwriting jitter, relying solely on subjective human observation. Determination of grade primarily relies on the physician's expert experience, and there is no uniform standard across different physicians.
[0005] 2. High medical pressure: During the first three to four weeks of CAR-T cell infusion, intensive dynamic monitoring is required, at least twice a day, and in extreme cases, four times a day, every two hours. The evaluation of a single patient can take nearly half an hour. The patient's handwriting is the core step in the ICANS assessment, relying on the experience and judgment of professional doctors. Summary of the Invention
[0006] The purpose of this invention is to provide a data toxicity assessment method and system based on font analysis. By collecting information using a pen-tip sensor, it can effectively determine a patient's control over their hand nerves. The AI model can automatically output the patient's handwriting ICE score.
[0007] In order to achieve the above objectives, the present invention provides a method for evaluating neurotoxicity based on font analysis, comprising the following steps:
[0008] S1, data acquisition, recording the patient's handwriting process to obtain multimodal writing process data, including the number of pauses, the number of releases, the stable value of the pen tip pressure, the longest straight line length, the font tilt angle, heart rate, and blood pressure;
[0009] S2, automatic feature engineering, analyzes the collected multimodal writing data, extracts multiple sets of feature data sets including reaction time series features, spatial topological features, and dynamic biomechanical parameters, and uses a hybrid screening strategy to evaluate the value of each feature, screening out valuable feature data sets, and thus forming a feature system for evaluating the patient's writing process;
[0010] S3, model training, using linear models and / or decision tree models to construct an ICE score to assess the severity of neurotoxicity in patients;
[0011] S4, model application, automatically generates a model business effect report, which shows the model's performance indicators, prediction accuracy, and performance in actual applications. The evaluation results are input into the automatic feature engineering step to re-screen the feature data set.
[0012] Furthermore, the number of pauses is that the writing image does not change for more than T time, which is considered as one pause.
[0013]
[0014] where t i is the time when the pixel change amplitude exceeds two pictures, and I is the number of pauses.
[0015] Furthermore, the number of times of releasing the hand is the number of times the pressure received by the pen tip is less than F.
[0016]
[0017] Among them, P i The pen pressure value collected at time i.
[0018] Furthermore, the T test results are calculated based on the two group means μP2 and μP2 of the pen tip pressure, the variances S1 and S2, and the sample sizes n1 and n2.
[0019]
[0020] Furthermore, the longest line segment with a curvature less than R is extracted from the image according to the recorded pen-fall coordinates, where the curvature is calculated as:
[0021]
[0022] in, and is a variable vector.
[0023] Furthermore, the calculation process of the font tilt angle includes:
[0024] Split the written text using CV and OCR technology;
[0025] Calculate the angle of each character from the central vertical line;
[0026] The mean, median, variance, and T-test results of the angles are used to describe the font tilt angle.
[0027] Furthermore, the multimodal writing data is extracted according to different acquisition parameters to form a collection of multiple groups of writing data. The value of each feature is evaluated by WOE, IV, and slot-wise AUC analysis methods to screen out valuable features.
[0028] Furthermore, a linear regression model is used to predict the output of the feature combination of multimodal writing data:
[0029] y=w1x1+w2x2+…+w n x n +b
[0030] Among them, w is the weight parameter, x is the feature, and b is the bias term;
[0031] Logistic regression uses the sigmoid function to implement probability mapping, and then solves the optimal weight vector by reducing the value of the loss function.
[0032] Furthermore, a random forest-based model was used to output the severity level of neurotoxicity through the average prediction of multiple decision trees.
[0033] In another aspect, the present invention further provides a neurotoxicity level assessment system based on font analysis, comprising:
[0034] Data acquisition module 1 records the patient's handwriting process to obtain multimodal writing process data, including the number of pauses, the number of releases, the stable value of the pen tip pressure, the longest straight line length, the font tilt angle, heart rate, and blood pressure;
[0035] Automatic feature engineering module 2 analyzes the collected multimodal writing data to extract multiple sets of feature data sets including time series features, spatial topological features, and dynamic biomechanical parameters. A hybrid screening strategy is used to evaluate the value of each feature, screening out valuable feature data sets to form a feature system for evaluating the patient's writing process.
[0036] Model training module 3, using linear models and / or decision tree models to construct a score to assess the severity of neurotoxicity in patients;
[0037] Model application module 4 automatically generates a model business effect report, which shows the model's performance indicators, prediction accuracy, and performance in actual applications, and inputs the evaluation results into the automatic feature engineering step to re-screen the feature data set.
[0038] This invention provides a data toxicity level assessment method and system based on font analysis. Through automatic feature processing and feature derivation of patient handwriting data, an effective feature system can be quickly constructed, providing rich input information for machine learning models. The collected data is used to score and output the patient's handwriting using ICE. The ICE score effectively reflects the severity of the patient's neurotoxicity. Based on the interactive approach of a large AI model, this invention can completely replace the complex work of nurses. Combined with periodic data, it can also provide patients with diagnostic trend analysis, helping to provide early warning of risks. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. 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 paying any creative work.
[0040] Figure 1 This is a flow chart of a method for evaluating neurotoxicity level based on font analysis according to an embodiment of the present invention.
[0041] Figure 2 This is a system framework diagram of a neurotoxicity level assessment system based on font analysis according to an embodiment of the present invention. DETAILED DESCRIPTION
[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.
[0043] In the following description, reference is made to “some embodiments”, which describes a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.
[0044] If similar descriptions of "first / second" appear in the application documents, the following explanation is added. In the following description, the terms "first\second\third" involved are merely used to distinguish similar objects and do not represent a specific order for the objects. It can be understood that "first\second\third" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.
[0045] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein are for the purpose of describing the embodiments of this application only and are not intended to limit this application.
[0046] The present invention provides a method and system for evaluating neurotoxicity levels based on font analysis, which significantly improves the efficiency of neurotoxicity level assessment through a hardware and software integrated solution. By collecting information with the help of a "pen tip" sensor, the patient's control over the hand nerves can be effectively judged. The AI model can automatically output the patient's handwriting ICE score and record the patient's writing process through the Pad panel, providing convenience for doctors to effectively trace back and evaluate. At the same time, the interactive method based on the AI large model can completely replace the complicated work of nurses. Combined with periodic data, it can also provide patients with diagnostic trend analysis to help early warning of risks. In addition, the RAG solution based on the large model can provide doctors with professional auxiliary treatment suggestions. As the collected data continues to increase, the model effect will continue to be optimized, further saving labor costs.
[0047] The following describes a method and system for evaluating neurotoxicity levels based on font analysis according to an embodiment of the present invention with reference to the accompanying drawings. First, a method for evaluating neurotoxicity levels based on font analysis according to an embodiment of the present invention will be described with reference to the accompanying drawings.
[0048] Example 1
[0049] Figure 1 FIG. 1 is a flow chart of a method for evaluating neurotoxicity level based on font analysis according to an embodiment of the present invention. Figure 1 As shown in Figure 1, this evaluation method includes the following steps: S1 data collection, S2 automatic feature engineering, S3 model training, and S4 model application. After the model is launched, new data from the business application production process is continuously fed back to the model training system in a closed loop, allowing the model to self-learn, adapt, and self-optimize. This forms a closed-loop feedback mechanism, automatically iterating and updating, preventing model failure and automatically updating without manual intervention.
[0050] In this embodiment, the data collection in step S1 mainly relies on the "pen tip" sensor and iPad. The "pen tip" writes on the iPad, and the "pen tip" collects the patient's heart rate, blood pressure and other physical data as well as pen tip pressure, pen grip, pen amplitude and vibration frequency through millisecond-level sensors. The iPad records the entire writing process and images to form the patient's writing process data.
[0051] In this embodiment, step S2 is to derive and screen features based on the patient's writing process data and the result data of whether or not the patient suffers from ICANS. For example, when the patient's writing process data is further processed, the number of pauses and the duration of the pauses in the writing process will be carefully analyzed based on the writing records on the iPad. By considering these parameters, we can better understand the patient's thinking coherence and concentration during the writing process. At the same time, the number of times the pressure drops above a certain threshold is analyzed based on the pen pressure record, and it is defined as the number of times the hand is released. This indicator can reflect the patient's hand nerve control ability and fatigue level. In addition, the image formed by the writing on the iPad is analyzed to find the longest line segment length of the straight line writing, etc. These image analyses can provide information about the patient's writing stability and fine motor control.
[0052] Specifically, the feature calculation process of automatic feature engineering is as follows:
[0053] The number of pauses is when the writing image does not change for more than T time, which is considered a pause.
[0054]
[0055] Where t_i is the time when the pixel change amplitude exceeds two pictures, and I is the number of pauses.
[0056] The number of times you release your hand is the number of times the pressure on the pen tip is less than F.
[0057]
[0058] Among them, P i The pen pressure value collected at time i.
[0059] Stability of pen tip pressure: The T-test results are calculated by the two groups of mean μP2 and μP2 of pen tip pressure, variance S1, S2, sample size n1, n2,
[0060]
[0061] The longest writing length is obtained by extracting the longest line segment with a curvature less than R from the image based on the pen coordinates recorded by the iPad.
[0062] According to the recorded pen-fall coordinates, the longest line segment with a curvature less than R is extracted from the image, where the curvature is calculated as:
[0063]
[0064] in, and is a variable vector.
[0065] Furthermore, the calculation process of the font tilt angle includes:
[0066] Split the written text using CV and OCR technology;
[0067] Calculate the angle of each character from the central vertical line;
[0068] The mean, median, variance, and T-test results of the angles are used to describe the font tilt angle.
[0069] Specifically, during the feature collection process, the collected multimodal writing data is analyzed to extract multiple feature datasets, including reaction time series features, spatial topological features, and dynamic biomechanical parameters. For example, the number of pauses can be sampled at intervals of 5 seconds, 10 seconds, and 30 seconds, forming multiple sampling groups. The curvature can be sampled at multiple angles. Based on these multiple feature sampling groups, multiple feature datasets are generated.
[0070] The acquired feature dataset was analyzed for importance using WOE (Weight of Evidence), IV (Information Value), and slot-wise AUC (Area Under the Curve) to identify valuable features, thereby forming a feature system for the patient's writing process. This feature system will provide a solid foundation for subsequent machine learning model training.
[0071] Specifically, WOE uses an equidistant binning method to bin each set of traits separately, and calculates the WOE value of each bin for each set of features:
[0072]
[0073] The confirmed ratio is the ratio of confirmed neurotoxicity samples in the current bin to all confirmed samples. The unconfirmed ratio is the ratio of unconfirmed samples in the current bin to all unconfirmed samples.
[0074] Calculate the IV value for each feature set to quantify its importance:
[0075] IV = ∑ (undiagnosed proportion i - diagnosed proportion i) × WOEi
[0076] Divide the features such as heart rate, blood pressure, pen tip pressure, pen grip strength, pen amplitude into independent feature groups.
[0077] For example, the physiological group includes heart rate and blood pressure, and the behavioral group includes pen tip pressure, grip strength, and amplitude. AUC values are calculated for each feature within each feature group, and low-contribution features (e.g., AUC < 0.6) are removed. For the retained feature groups, the overall AUC is calculated for each group and sorted by AUC, prioritizing the high-contribution groups.
[0078] In this embodiment, step S3 involves training using a machine learning model. This module attempts to construct tree models and linear models to adapt to different data characteristics and problem types. Simultaneously, automatic parameter tuning algorithms such as zero-order optimization and grid-search are applied to optimize model parameters and enhance model performance. These automatic parameter tuning algorithms can quickly find the optimal solution among a large number of parameter combinations, improving model accuracy and generalization capabilities.
[0079] Specifically, the linear regression model predicts the output of the feature combination of multimodal writing data:
[0080] y=w1x1+w2x2+…+w n x n +b
[0081] Among them, w is the weight parameter, x is the feature, and b is the bias term;
[0082] Logistic regression uses the sigmoid function to implement probability mapping, and then solves the optimal weight vector by reducing the value of the loss function.
[0083] Then, a random forest model was used to average the predictions of multiple decision trees to output the severity level of neurotoxicity. The scoring method used was to construct a score where the child nodes of the decision trees output the probability of disease, which could be directly used as the risk score.
[0084] Example 2
[0085] In another aspect, the present invention further provides a neurotoxicity level assessment system based on font analysis, comprising:
[0086] Data acquisition module 1 records the patient's handwriting process to obtain multimodal writing process data, including the number of pauses, the number of releases, the stable value of the pen tip pressure, the longest straight line length, the font tilt angle, heart rate, and blood pressure;
[0087] Automatic feature engineering module 2 analyzes the collected multimodal writing data to extract multiple sets of feature data sets including time series features, spatial topological features, and dynamic biomechanical parameters. A hybrid screening strategy is used to evaluate the value of each feature, screening out valuable feature data sets to form a feature system for evaluating the patient's writing process.
[0088] Model training module 3, using linear models and / or decision tree models to construct a score to assess the severity of neurotoxicity in patients;
[0089] Model application module 4 automatically generates a model business effect report, which shows the model's performance indicators, prediction accuracy, and performance in actual applications, and inputs the evaluation results into the automatic feature engineering step to re-screen the feature data set.
[0090] Although the embodiments of the present invention have been shown and described above, it should be understood that the above embodiments are illustrative and are not to be construed as limiting the present invention. Those skilled in the art may make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention. Any modifications, equivalent substitutions, improvements, and the like made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.
Claims
1. A neurotoxicity level assessment method based on font analysis, characterized in that: The following steps are involved: S1, data acquisition, recording the patient's handwriting process to obtain multimodal writing process data, including the number of pauses, the number of releases, the stable value of the pen tip pressure, the longest straight line length, the font tilt angle, heart rate, and blood pressure; S2, automatic feature engineering, analyzes the collected multimodal writing data, extracts multiple sets of feature data sets including reaction time series features, spatial topological features, and dynamic biomechanical parameters, and uses a hybrid screening strategy to evaluate the value of each feature, screening out valuable feature data sets, and thus forming a feature system for evaluating the patient's writing process; S3, model training, using linear models and / or decision tree models to construct an ICE score to assess the severity of neurotoxicity in patients; S4, model application, automatically generates a model business effect report, which shows the model's performance indicators, prediction accuracy, and performance in actual applications. The evaluation results are input into the automatic feature engineering step to re-screen the feature data set.
2. The neurotoxicity level assessment method based on font analysis according to claim 1, characterized in that: The number of pauses is considered to be one pause if the writing image does not change for more than T time. where t i is the time when the pixel change amplitude exceeds two pictures, and I is the number of pauses.
3. The neurotoxicity level assessment method based on font analysis according to claim 1, characterized in that: The number of times of releasing the hand is the number of times the pressure received by the pen tip is less than F. Among them, P i The pen pressure value collected at time i.
4. The neurotoxicity level assessment method based on font analysis according to claim 1, characterized in that: The T test results are calculated by the two groups of mean μP2 and μP2 of the pen tip pressure, variance S1, S2, sample size n1, n2, 5. The neurotoxicity level assessment method based on font analysis according to claim 1, characterized in that: According to the recorded pen-fall coordinates, the longest line segment with a curvature less than R is extracted from the image, where the curvature is calculated as: in, and is a variable vector.
6. The neurotoxicity level assessment method based on font analysis according to claim 1, characterized in that: The calculation process of the font tilt angle includes: Split the written text using CV and OCR technology; Calculate the angle of each character from the central vertical line; The mean, median, variance, and T-test results of the angles are used to describe the font tilt angle.
7. The neurotoxicity level assessment method based on font analysis according to any one of claims 1 to 6, characterized in that: The multimodal writing data is extracted according to different acquisition parameters to form a set of multiple groups of writing data. The value of each feature is evaluated by WOE, IV, and slot-wise AUC analysis methods to screen out valuable features.
8. The neurotoxicity level assessment method based on font analysis according to any one of claims 1 to 7, characterized in that: The linear regression model is used to predict the output of the feature combination of multimodal writing data: y=w1x1+w2x2+…+w n x n +b Among them, w is the weight parameter, x is the feature, and b is the bias term; Logistic regression uses the sigmoid function to implement probability mapping, and then solves the optimal weight vector by reducing the value of the loss function.
9. The neurotoxicity level assessment method based on font analysis according to any one of claims 1 to 7, characterized in that: A random forest-based model was used to output the ICE score of the severity of neurotoxicity through the average prediction of multiple decision trees.
10. A neurotoxicity level assessment system based on font analysis, characterized in that: include: Data acquisition module 1 records the patient's handwriting process to obtain multimodal writing process data, including the number of pauses, the number of releases, the stable value of the pen tip pressure, the longest straight line length, the font tilt angle, heart rate, and blood pressure; Automatic feature engineering module 2 analyzes the collected multimodal writing data, extracts multiple sets of feature data sets including reaction time series features, spatial topological features, and dynamic biomechanical parameters, and uses a hybrid screening strategy to evaluate the value of each feature, screening out valuable feature data sets to form a feature system for evaluating the patient's writing process. Model training module 3, using linear models and / or decision tree models to construct a score to assess the severity of neurotoxicity in patients; Model application module 4 automatically generates a model business effect report, which displays the model's performance indicators, prediction accuracy, and performance in actual applications, and inputs the evaluation results into the automatic feature engineering step to re-screen the feature data set.