A method, system and device for assisting in the evaluation of the oral motor function of a newborn

Abstract

The present application belongs to the technical field of medical health, and discloses a new-born oral cavity movement function auxiliary evaluation method, system and device, which comprises the following steps: acquiring a non-nutritive sucking pressure signal of a new-born and preprocessing the signal; intercepting an effective data segment through a multi-strategy analysis window; identifying effective positive pressure peaks and negative pressure peaks by using a local extremum detection method combined with an amplitude threshold, a minimum peak spacing and a signal quality constraint; determining sucking bursts, pauses and isolated sucking events according to the time interval of adjacent positive pressure peaks; extracting multi-dimensional NNS kinetic characteristics, eliminating redundancy in the training data through correlation analysis, and obtaining a non-redundant feature subset; inputting the subset and clinical features into a machine learning model with a regularization constraint, training the model with a clinical evaluation scale as a label, and finally outputting auxiliary evaluation probability, risk level and signal quality information. The present application can objectively and quantitatively analyze NNS signals, and provide auxiliary reference for clinical function evaluation and feeding management.

Description

Technical Field

[0001] This invention belongs to the field of medical and health technology, and specifically relates to a method, system and device for assisting in the assessment of oral motor function in newborns. Background Technology

[0002] Newborns, especially premature infants, low birth weight infants, and newborns with neurological or respiratory diseases, may exhibit oral motor function problems such as unstable sucking rhythm, insufficient sucking strength, and poor organization of sucking movements. Timely and objective identification of these problems helps clinicians develop individualized feeding training and care plans.

[0003] Currently, the main clinical methods for assessing the oral motor development of newborns include: (1) Observation method: relying on the subjective judgment of medical staff on sucking actions, lacking objectivity and consistency, and low inter-assessment reliability; (2) Standardized scale method: such as the Newborn Oral Motor Assessment Scale (NOMAS), which requires professional training, the assessment results are affected by the observer's experience, and cannot be tracked dynamically in real time; (3) Nutritional sucking measurement method: indirectly assessed through weight gain during breastfeeding, with significant individual differences; (4) Instrument method: using specialized sucking measurement equipment to record oral pressure, which can provide quantitative data, but the equipment is expensive, the operation is complicated, and existing methods mostly remain at the level of single indicator description, lacking a systematic prediction model.

[0004] Non-nutritive sucking (NNS) refers to sucking activities performed by newborns without nutritional intake. Information such as sucking pressure, sucking frequency, rhythm variability, number of sucking bursts, duration of sucking bursts, and pause structures in NNS pressure signals can be used to characterize the organization and coordination of oral movements related to non-nutritive sucking in newborns.

[0005] However, existing research on neonatal oral motor function (NNS) largely remains at the level of descriptive statistical analysis, lacking systematic machine learning predictive models, making it difficult to translate NNS dynamic parameters into clinical decision support tools. Furthermore, existing technologies, such as CN202411157117, disclose a weighted formula-based method for assessing sucking ability index. This method uses a fixed analytical formula to calculate continuous scores, without addressing the identification of sucking bursts and pauses, nor employing machine learning classification models. Essentially, this method is a rating scale that cannot output disease prediction probabilities, and its formula parameters lack physiological basis and experimental validation. Therefore, there is an urgent need for an objective assessment method for neonatal oral motor function. Summary of the Invention

[0006] The present invention aims to at least partially solve the aforementioned technical problems. Therefore, the objective of the present invention is to provide a method, system, and device for assessing oral motor function in newborns. The present invention can automatically extract multi-dimensional dynamic features from non-nutritive sucking pressure signals and achieve objective and quantitative assessment of oral motor function through machine learning algorithms.

[0007] To achieve the above objectives, the present invention provides a method for assessing oral motor function in newborns, the method comprising:

[0008] Raw pressure signal data of non-nutritive sucking (NNS) in newborns are acquired, and the raw data is preprocessed to obtain preprocessed data that meets the analysis requirements.

[0009] A multi-strategy analysis window capture method is used to determine at least one valid analysis window of 2 consecutive minutes from the preprocessed data. The multi-strategy analysis window capture method includes a fixed window mode and a multi-window integrated mode.

[0010] The pressure signal within the analysis window is subjected to suction peak identification. Positive pressure peaks are identified using a local maximum detection method, and negative pressure peaks are identified using a local minimum detection method. Noise and spurious peaks are eliminated by combining amplitude threshold, minimum peak spacing, and signal quality constraints. The amplitude threshold for positive pressure peaks is preferably 1 cmH2O, and the minimum peak spacing is preferably 300 ms.

[0011] Suckling events are structurally determined based on the time interval between adjacent effective positron pressure peaks. A continuous positron pressure peak sequence containing at least two effective positron pressure peaks and with an interval between adjacent positron pressure peaks of less than 1000 ms is identified as a sucking burst. A single effective positron pressure peak that is not classified as a sucking burst is identified as an isolated sucking event, and the pause period between adjacent sucking bursts is identified.

[0012] Based on the sucking peaks and sucking bursts, multi-dimensional NNS dynamic characteristic indicators are extracted, and the redundancy of the characteristic indicators is eliminated by correlation analysis to obtain a non-redundant feature subset.

[0013] The non-redundant feature subset and neonatal clinical baseline features are used as inputs and fed into a pre-trained machine learning classification model with regularization constraints for processing, so as to output auxiliary assessment results of neonatal non-nutritive sucking-related oral motor functions.

[0014] Optionally, the preprocessing includes: converting the pressure unit from mmHg to cmH2O (conversion factor 1.36); identifying abnormal pressure values ​​based on the effective range of the pressure sensor and the preset physiological effective range; correcting isolated abnormal points with short durations using local median, mean of adjacent effective values, or other local statistical methods; marking continuously occurring abnormal data segments with a duration exceeding a preset threshold as invalid data segments and not identifying sucking events; completing short-term missing data with a duration not exceeding the preset threshold using linear interpolation; marking long-duration missing data segments as invalid data segments; performing baseline drift correction on the pressure signal to reduce low-frequency drift caused by sensor temperature changes, nipple position changes, and long-term acquisition; determining filtering parameters based on the actual sampling frequency and denoising the signal using moving average filtering, low-pass filtering, or other smoothing methods. In one specific embodiment, when the sampling frequency is 100Hz, a moving average filter with a window width of 5 sampling points can be used. When the sampling frequency changes, the filter window width is adjusted according to the actual sampling frequency to ensure that the filtering time scale is basically consistent under different acquisition devices. Optionally, the preprocessing also includes evaluating the signal quality. The signal quality evaluation includes at least one or more of the following: percentage of valid data, percentage of missing data, percentage of outlier data, signal saturation, baseline drift, and noise level. Data segments that do not meet the preset signal quality requirements are not considered valid analysis windows, or are marked as low-confidence results in the evaluation results.

[0015] Optionally, in the fixed window mode, after completing sensor stability and signal quality checks, the first consecutive 2-minute data segment that meets the effective data ratio requirement is selected as the analysis window. This mode is suitable for standardized short-term recording scenarios in clinical settings and can reduce selection bias caused by the operator's subjective selection of signal segments. Optionally, in the multi-window integration mode, multiple consecutive 2-minute analysis windows are extracted from longer non-nutritive sucking records, and the dynamic characteristics of each analysis window are calculated separately. For multiple effective analysis windows, the arithmetic mean, weighted mean, median, or truncated mean can be used to integrate similar features of each window. Specifically, when using non-overlapping windows, the step size between adjacent windows is not less than 2 minutes; when using overlapping windows, the feature values ​​are weighted according to the window overlap ratio.

[0016] Optionally, the peak identification also includes a local extremum localization rule. The local extremum detection method further includes a peak localization rule: that is, within a preset local neighborhood, data points with data values ​​greater than other sampling points in the neighborhood are identified as positive pressure candidate peaks, and data points with data values ​​less than other sampling points in the neighborhood are identified as negative pressure candidate peaks. The positive pressure peak amplitude refers to the positive pressure change amplitude of the positive pressure candidate peak relative to the zero pressure baseline or local baseline. Only by combining the positive pressure amplitude threshold (A≥1cmH2O), the preset negative pressure threshold, and the minimum peak spacing (≥300ms) constraint can it be identified as a valid positive pressure peak, resulting in a positive pressure peak set P+ and a negative pressure peak set P-, and reducing the identification of false peaks caused by noise.

[0017] Optionally, the redundancy elimination step is as follows: calculate the Pearson correlation coefficient or Spearman correlation coefficient between each pair of feature indicators; set the correlation threshold to |r|>0.85; for feature pairs whose absolute correlation coefficient is greater than or equal to the threshold, filter them according to the preset feature retention rules to generate the non-redundant feature subset.

[0018] Optionally, the preset feature retention rule includes at least one of the following:

[0019] (a) For feature pairs with mathematical complementarity, retain the one with better clinical interpretability; (b) For feature pairs with high positive correlation, retain the one that contains more comprehensive sucking structure information, has higher repeated measure stability, or makes a greater contribution to model prediction.

[0020] (c) For highly correlated pairs of clinical baseline features, retain at least one of the following factors: clinical importance, repeated measures stability, and model prediction contribution;

[0021] (d) Features that are not relevant enough to be removed but have independent clinical significance shall be retained.

[0022] Optionally, the machine learning classification model with added regularization constraints includes at least one of the following: a logistic regression (LR) model with L2 regularization, a support vector machine (SVM) model using a radial basis function (RBF) kernel function with L2 regularization, a random forest (RF) model with a maximum tree depth limit, or an extreme gradient boosting tree (XGBoost) model with added L1 and L2 regularization terms. The model parameters and regularization parameters of the machine learning classification model are selected through inner-layer cross-validation. The classification labels of the machine learning classification model are determined by preset clinical reference standards, including the assessment conclusions of the Neonatal Oral Motor Assessment Scale (NOMAS) or the assessment conclusions of the Preterm Infant Oral Feeding Readiness Assessment Scale (PIOFRA).

[0023] Optionally, the method further includes: evaluating the machine learning classification model using hierarchical five-fold cross-validation, independently calculating and reporting the area under the ROC curve, accuracy, recall (as an important indicator reflecting the ability to detect anomalies), specificity, and F1 score; and using the DeLong test to assess whether the difference in the area under the ROC curve (AUC) between different models is statistically significant (e.g., significance level α = 0.05).

[0024] This invention also provides a neonatal oral motor function assessment system, comprising: a data acquisition module, a preprocessing module, a multi-strategy window truncation module, a peak identification module, a sucking burst determination and feature extraction module, a feature processing module, a regularized assessment modeling module, and an assessment result output module. The data acquisition module acquires raw pressure signal data from non-nutritive sucking in newborns; the preprocessing module performs preprocessing steps; the multi-strategy window truncation module performs window truncation methods; the peak identification module performs local extremum detection to identify effective positive and effective negative pressure peaks in the pressure signal; the sucking burst determination and feature extraction module identifies the sucking burst structure and extracts NNS dynamic features; the feature processing module performs correlation analysis and redundancy elimination steps; the regularized assessment modeling module performs machine learning modeling, parameter tuning, probability calibration, and model evaluation; and the assessment result output module provides auxiliary assessment probabilities, risk levels, or signal quality prompts based on the model output. All modules work collaboratively to achieve the methods described above.

[0025] The present invention also provides an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor reads and executes the computer program, it implements the auxiliary evaluation method described in any of the preceding claims.

[0026] The beneficial effects of this invention are as follows:

[0027] This invention automates the data processing, feature extraction, and modeling analysis of non-nutritive sucking pressure signals in newborns. It replaces the subjective judgments in traditional observation and scale methods with objective and quantitative indicators, reducing subjective differences among assessors and improving the consistency of results.

[0028] This invention not only analyzes basic parameters such as the frequency and intensity of sucking, but also introduces a refined analysis of the "sucking burst-pause" structure, and constructs a feature system that includes multiple dimensions such as counting, proportion, structure, frequency, and intensity, which can more comprehensively characterize the rhythmicity, organization, and coordination of newborns' non-nutritive sucking activities.

[0029] This invention introduces regularized machine learning models into NNS data analysis, achieving a leap from descriptive statistics to predictive evaluation. Overfitting is effectively suppressed through strategies such as L1 / L2 regularization and limiting tree depth. Rigorous evaluation using methods such as cross-validation and DeLong's test ensures the model's good generalization ability and predictive stability on unknown data.

[0030] The multi-strategy analysis window truncation method of the present invention can ensure standardized clinical operation through a fixed truncation mode, and adapt to long-term recording through a mean integration mode, thereby improving the global representativeness of features. This allows the method to flexibly cope with different clinical data collection scenarios and enhances its robustness. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the NNS signal processing and feature extraction process in the method of the present invention;

[0032] Figure 2 This is a schematic diagram of the machine learning evaluation and modeling process in the method of the present invention;

[0033] Figure 3 This is a schematic diagram of the neonatal oral motor function assessment system of the present invention;

[0034] Figure 4 A schematic diagram comparing the ROC curves of four regularized machine learning models;

[0035] Figure 5 This is a schematic diagram of the hardware structure of the electronic device in an embodiment of the present invention. Detailed Implementation

[0036] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0037] like Figure 1 As shown, the initial step of the neonatal oral motor function assessment method of the present invention is to acquire the non-nutritive sucking pressure signal of the neonate. Specifically, a miniature pressure sensor installed inside the pacifier can be used for acquisition. The sensor's range needs to cover -200 to +200 cmH2O, and the sampling frequency should not be less than 100 Hz. The acquired data is recorded in a time (ms)-pressure (mmHg) time series format and transmitted to a terminal device for subsequent processing via wired or wireless means.

[0038] The next step is to preprocess the raw data. In one specific embodiment, this step includes:

[0039] (1) Unit conversion: The pressure unit is converted from mmHg to cmH2O, which is more commonly used in clinical practice. The conversion formula is P(cmH2O)=P(mmHg)×1.36.

[0040] (2) Outlier handling: detect and handle extreme pressure values ​​that exceed the preset physiological range, such as positive pressure greater than 200cmH2O or negative pressure less than -200cmH2O. For isolated outlier points, use the local median or Hampel method to handle them; when a continuous outlier segment exceeds the threshold, the segment is marked as invalid.

[0041] (3) Missing value completion: For missing sampling points generated during data transmission or acquisition, linear interpolation is used to fill them in to ensure the continuity of the signal; long missing segments are marked as invalid.

[0042] (4) Baseline drift correction: Zero-point calibration and baseline drift correction are performed on the pressure signal to reduce the low-frequency drift caused by sensor temperature changes, nipple position changes and long-term acquisition, so that the corrected zero pressure baseline is kept near 0cmH2O.

[0043] (5) Noise filtering: Determine the filtering parameters based on the actual sampling frequency, and use moving average low-pass filtering or other low-pass filtering methods to remove noise.

[0044] (6) Signal quality evaluation: The signal quality evaluation includes at least one of the following: percentage of valid data, percentage of missing data, percentage of abnormal data, signal saturation, baseline drift, and noise level. Data segments that do not meet the preset signal quality requirements are not considered valid analysis windows, or are marked as low-confidence data in subsequent evaluation results.

[0045] After preprocessing, the multi-strategy analysis window capture step is performed. This invention provides two modes for users to choose from according to clinical scenarios; mode A is used by default.

[0046] Mode A (Fixed Truncation): The first consecutive 120-second valid data segment after quality inspection is selected as the analysis window W. This mode is suitable for standardized short-term records and can reduce selection bias introduced by the operator's subjective selection of the best segment.

[0047] Mode B (Multi-Window Mode): When the recording duration is long (e.g., exceeding 10 minutes), the system may prompt the user to switch to this mode. Multiple consecutive 2-minute analysis windows are extracted with a default step size of 120 seconds. When using overlapping windows, analysis windows are extracted with a preset step size of less than 120 seconds, and weights are adjusted according to the window overlap ratio. Features are extracted separately from each analysis window that passes signal quality control, and then integrated using the median, truncated mean, or weighted mean. This mode reduces the impact of a single abnormal window on the final result and more comprehensively reflects the newborn's sucking behavior throughout the recording period. This mode better reflects the newborn's global sucking behavior throughout the recording period.

[0048] The suction peak was identified on the captured 2-minute window of data. Please continue reading. Figure 1 This invention employs a local maximum detection method that combines dual constraints:

[0049] (1) Amplitude threshold constraint: The amplitude of a valid compression event must satisfy A≥1cmH2O (equivalent to A_mmHg≥0.735mmHg). This can effectively filter out small fluctuations and noise in the signal baseline. For negative pressure peaks, the candidate peaks are required to be below the zero pressure baseline and their negative amplitude absolute value must reach the preset negative pressure threshold.

[0050] (2) Minimum peak spacing constraint: The time interval between two adjacent positive pressure peaks must be greater than 300ms. In the specific implementation, if the sampling rate is 100Hz, it is equivalent to setting the minimum span between adjacent maxima in the peak search algorithm to 30 sampling points. This constraint aims to prevent a single sucking action from being incorrectly identified as two or more adjacent pseudo-double peaks due to noise or irregular shape.

[0051] (3) Peak location rule: Under the premise of satisfying the above constraints, the signal is traversed. Positive pressure peaks are detected by local maximum value, and negative pressure peaks are detected by local minimum value or by taking the negative of the signal and then detecting the maximum value. Through this method, the set of positive pressure peaks P+={p1,p2,...,p n} and its corresponding time sequence T+={t1,t2,...,t n}, and the negative pressure peak set P-={q1,q2,...,q m} and its corresponding time sequence T-={s1,s2,...,s m}

[0052] After identifying the peaks, the sucking bursts and pauses are determined. This invention defines:

[0053] Sucking burst: A sucking burst must contain at least two valid positive pressure peaks, and the time interval between any two adjacent positive pressure peaks must be less than 1000ms (Δt=t_{i+1}-t_i<1000ms).

[0054] Pause: The resting time interval between two consecutive sucking bursts, with a duration greater than or equal to 1000 ms.

[0055] Isolated Suck: A single effective positive pressure peak that is not included in any sucker burst sequence.

[0056] Based on the above identification results, multi-dimensional NNS dynamic characteristic indicators are extracted. It is important to emphasize that since all raw pressure data has been uniformly converted to cmH2O units in the preprocessing step, all calculations involving pressure, amplitude, or peak value in the table below use cmH2O units for both input and output values ​​by default, requiring no further unit conversion. The table below lists some core indicators and their calculation definitions:

[0057] Indicator Categories Indicator Name Calculation definition Counting Total oral motion events (TOME) The total number of effective positive pressure peaks within the analysis window is N = |P + | Counting Sucking-related NNS period count The total number of barotropic peaks belonging to a burst sequence = Σ|Bk| Counting Number of isolated events TOME-Sucking-related NNS cycle number Proportional NNS sucking burst cycle percentage (%) Sucking-related NNS cycle count / TOME × 100 Sudden structural class NNS sucking bursts per minute Analyze the total number of sucking bursts K within the analysis window / window duration (min). Sudden structural class Average number of cycles per sucking burst (1 / K)×Σ|Bk|, mean number of suckings within each sucking burst Sudden structural class Mean duration of the sudden event (s) The average time difference between the last peak and the first peak of each complete sucking burst. Frequency intensity Burst frequency mean (Hz) Calculate the average sucking frequency within each complete sucking burst and then take the average value. Frequency intensity <![CDATA[Mean burst amplitude (cmH2O)]]> After calculating the average amplitude of the positive pressure peak in each complete suction burst, the average value was taken. Frequency intensity Total cycles per minute (times / min) TOME / Window Duration (min) Pressure intensity <![CDATA[Maximum Positive Pressure Peak (cmH2O)]]> max(A_i), i∈P+ Pressure intensity <![CDATA[Absolute value of peak sucking negative pressure (cmH2O)]]> |min(A_j)|,j∈P-

[0058] The extracted NNS dynamic features are integrated with clinical baseline features (e.g., gestational age, birth weight, days of birth, and whether a nasogastric tube is in place) to construct an initial feature matrix. Subsequently, a feature redundancy elimination step is performed. In one specific embodiment, this step is as follows:

[0059] (1) Calculate the Pearson correlation coefficient r between all feature pairs in the initial feature matrix and generate a correlation heatmap.

[0060] (2) Set the high correlation threshold as |r|>0.85.

[0061] (3) Eliminate highly relevant feature pairs according to preset rules:

[0062] For features with a definite mathematical complementary relationship, such as "NNS outbreak cycle proportion" and "isolated event proportion", the judgment is no longer based solely on the sample correlation coefficient, but rather on retaining the one with stronger clinical interpretability, such as retaining the NNS outbreak cycle proportion.

[0063] For feature pairs that have a certain negative correlation but are clinically independent and whose correlation coefficients do not reach the removal threshold, such as the correlation coefficient of "mean sudden amplitude" and "mean sudden frequency" in a confirmatory experiment being approximately -0.71, both will be retained.

[0064] For clinical baseline characteristics such as gestational age and birth weight, not only are they screened based on correlation coefficients, but also their clinical significance, repeated measures stability, and model predictive contribution are considered to determine which characteristics to retain.

[0065] Finally, an optimal non-redundant feature subset is obtained for subsequent modeling.

[0066] Before model training, outcome labels for the samples are determined based on preset clinical reference standards. In one specific embodiment, the assessment conclusion of the Neonatal Oral Motor Assessment Scale (NOMAS) is used as the primary outcome label. Samples with an assessment conclusion of normal sucking pattern are labeled as the normal group, and samples with an assessment conclusion of sucking disorder or sucking dysfunction are labeled as the abnormal group. For preterm infant samples, the assessment conclusion of the Preterm Infant Oral Feeding Readiness Assessment Scale (PIOFRA) can also be used to divide the samples into normal and abnormal groups of oral feeding readiness according to preset scoring thresholds or clinical judgment criteria. The NOMAS or PIOFRA assessment is preferably performed by clinical evaluators who have received standardized training and are unaware of the model output results; when the conclusions of two evaluators are inconsistent, a third evaluator reviews or makes a decision.

[0067] The sample exclusion criteria include: individuals with congenital structural abnormalities of the oral and maxillofacial region; individuals with severe motion artifacts, sensor detachment, continuous long-term signal loss, or signal quality that does not meet the preset requirements during the recording of non-nutritive sucking pressure signals.

[0068] Please see Figure 2 The non-redundant feature subset and the selected clinical baseline features are used as input and fed into a machine learning classification model with regularization constraints. Regularization is a key step in suppressing model overfitting and improving generalization ability. The present invention can employ the following models:

[0069] (1) Logistic Regression (LR+L2 Regularization): Search for the optimal regularization coefficient C in the range of [0.01,100] through cross-validation.

[0070] (2) Support Vector Machine (SVM+RBF kernel): The optimal penalty coefficient C (e.g. [0.01, 100]) and kernel parameter γ are determined by grid search.

[0071] (3) Random Forest (RF): Regularization is achieved by limiting the maximum tree depth (e.g., max_depth≤8) and increasing the minimum number of samples in the leaf nodes (e.g., min_samples_leaf=5).

[0072] (4) XGBoost: It directly introduces L1 regularization terms (reg_alpha) and L2 regularization terms (reg_lambda) into the model, and can limit the tree depth and learning rate, and optimize through cross-validation.

[0073] To objectively evaluate model performance, this invention employs a stratified five-fold cross-validation strategy. In each fold, AUC (Area Under the ROC Curve), precision, recall, and F1 score are calculated and recorded independently. Specifically, this invention emphasizes reporting precision and recall as two independent metrics because, in medical settings, the ability to detect anomalous samples (recall) is often as important as, or even more important than, overall classification accuracy (precision). Final performance is presented as the mean ± standard deviation of the five-fold cross-validation. Furthermore, the DeLong test is used to compare AUCs pairwise between different models to assess whether performance differences are statistically significant (e.g., p < 0.05).

[0074] In a confirmatory experiment involving 155 newborn samples, using the NOMAS assessment results as the primary reference standard, samples with normal sucking patterns were classified as the normal group, while samples with sucking disorders or sucking dysfunction were classified as the abnormal group. Samples with congenital oral and maxillofacial structural abnormalities, severe motion artifacts, or sensor detachment were excluded. Based on this, the performance of four regularized machine learning models was evaluated, and the results are shown in Table 1.

[0075] Table 1 Performance Evaluation of Four Regularized Machine Learning Models

[0076] Model Verify AUC Verification accuracy Validate recall Validate F1 LR+L2 regular expressions 0.770±0.050 0.671±0.074 0.743±0.081 0.754±0.074 Random Forest 0.757±0.075 0.677±0.068 0.712±0.085 0.707±0.072 SVM+RBF kernel 0.743±0.067 0.658±0.074 0.758±0.024 0.748±0.024 XGBoost+ Regularization 0.723±0.070 0.697±0.108 0.710±0.075 0.714±0.071

[0077] The results show that all four machine learning models have certain classification and prediction capabilities. Among them, the LR+L2 regularized model has the highest validation AUC (0.770±0.050) and the highest F1 score (0.754±0.074), demonstrating good generalization ability and stability in this validation experiment.

[0078] The paired comparison of AUC among the models showed that the differences in AUC among the four models did not reach a statistically significant level. The specific comparison method used should be determined based on the validation data structure; when there is an independent validation set, the DeLong test can be used; when cross-validation is used, the Bootstrap method based on out-of-bounds prediction probabilities is preferred.

[0079] Finally, the system classifies the risk of non-nutritive sucking-related oral motor function in newborns based on the probability P output by the selected model.

[0080] The classification threshold should be determined within the model training data based on preset clinical objectives, such as selecting an appropriate threshold while ensuring a certain recall rate. In a specific embodiment, the threshold can be set to 0.5, and the results can be output as follows:

[0081] P≥0.5: This indicates possible abnormalities in oral motor function, and further professional evaluation (such as the NOMAS scale) and timely oral motor intervention are recommended.

[0082] P<0.5: Oral motor function is considered normal, and continued routine observation is recommended.

[0083] Furthermore, this invention supports periodic dynamic monitoring, such as recommending repeating NNS measurements and assessments weekly or bi-weekly to dynamically track the developmental trajectory of neonatal oral motor function, providing an objective basis for clinical development and adjustment of individualized intervention programs.

[0084] Please see Figure 3 This invention also provides a neonatal oral motor function assessment system. The system includes: a data acquisition module, a preprocessing module, a multi-strategy window capture module, a peak recognition module, a sucking burst determination and feature extraction module, a feature processing module, a regularized assessment modeling module, and an assessment result output module. The functions of each module have been described in detail in the foregoing method; they work sequentially or collaboratively to jointly realize the assessment method of this invention.

[0085] Please see Figure 5 The method of the present invention can be implemented by an electronic device. This electronic device may include a processor, a memory, and a bus connecting the two. The memory stores a computer program, which, when executed by the processor, implements the evaluation method steps of the present invention.

[0086] The present invention also provides a non-transitory computer-readable storage medium, such as a CD, DVD, USB flash drive, solid-state drive, etc. This storage medium stores a computer software program containing instructions for implementing the method of the present invention. When the program is executed by a processor, it enables the implementation of a method for assessing oral motor function in newborns.

[0087] This invention is not limited to the above-described optional embodiments. Anyone can derive other various forms of products under the guidance of this invention. However, regardless of any changes made in their shape or structure, any technical solution that falls within the scope of the claims of this invention shall be protected by this invention.

Claims

1. A method for assisting in the assessment of oral motor function in newborns, characterized in that, The method includes: Raw data of non-nutritive sucking pressure signals from newborns are acquired, and the raw data is preprocessed to obtain preprocessed data. A multi-strategy analysis window interception method is used to determine at least one valid analysis window of data for a continuous 2 minutes from the preprocessed data. The multi-strategy analysis window interception method includes a fixed window mode and a multi-window integrated mode. The effective analysis window data is analyzed using a local extremum detection method that combines amplitude threshold, minimum peak spacing constraint, and signal quality constraint to identify sucker peaks. Specifically, local maximum detection is used for positive pressure peaks, and local minimum detection is used for negative pressure peaks. The amplitude threshold for positive pressure peaks is 1 cmH2O, and the minimum peak spacing is 300 ms. Suckling events are structurally determined based on the time interval between adjacent effective positive pressure peaks. A continuous positive pressure peak sequence containing at least two effective positive pressure peaks and with an interval between adjacent effective positive pressure peaks of less than 1000ms is identified as a sucking burst. Based on the sucking peaks and sucking bursts, multi-dimensional NNS dynamic characteristic indicators are extracted, and the redundancy of the characteristic indicators is eliminated by correlation analysis to obtain a non-redundant feature subset. The non-redundant feature subset and neonatal clinical baseline features are used as inputs and fed into a pre-defined machine learning classification model with regularization constraints for processing, so as to output auxiliary assessment results of neonatal non-nutritive sucking-related oral motor functions.

2. The method according to claim 1, characterized in that, The preprocessing includes: converting the pressure unit from mmHg to cmH2O; handling outliers for extreme pressure values ​​that exceed the physiological range or the effective range of the sensor; filling in short-term missing sampling points using linear interpolation; marking long-term missing data segments as invalid; performing baseline drift correction on the pressure signal; determining filtering parameters based on the actual sampling frequency; and using a moving average low-pass filter to denoise the signal.

3. The method according to claim 1, characterized in that, The fixed window mode extracts the first consecutive 2-minute valid data segment after passing the signal quality check as the analysis window; the multi-window integration mode extracts multiple consecutive 2-minute analysis windows from the longer NNS record, calculates feature indicators for each analysis window that passes the signal quality check, and then integrates the features using the arithmetic mean, weighted average, median, or truncated average; wherein, when using non-overlapping windows, the step size between adjacent windows is not less than 2 minutes; when using overlapping windows, the feature values ​​are weighted according to the window overlap ratio.

4. The method according to claim 1, characterized in that, The local extreme value detection method also includes peak positioning rules: within a preset local neighborhood, data points with data values ​​greater than other sampling points in the neighborhood are identified as positive pressure candidate peaks, and data points with data values ​​less than other sampling points in the neighborhood are identified as negative pressure candidate peaks; wherein, the positive pressure peak amplitude refers to the positive pressure change amplitude of the positive pressure candidate peak relative to the zero pressure baseline or local baseline, and it is only identified as a valid positive pressure peak when the positive pressure change amplitude is greater than or equal to 1cmH2O and meets the minimum peak spacing constraint.

5. The method according to claim 1, characterized in that, The redundancy elimination steps are as follows: calculate the Pearson correlation coefficient or Spearman correlation coefficient between each pair of feature indicators; set the correlation threshold to |r|>0.85; for feature pairs whose absolute correlation coefficient is greater than or equal to the threshold, filter them according to the preset feature retention rules to generate the non-redundant feature subset.

6. The method according to claim 5, characterized in that, The preset feature preservation rule includes at least one of the following: (a) For feature pairs with mathematical complementarity, retain the one with better clinical interpretability; (b) For feature pairs with high positive correlation, retain the one that contains more comprehensive sucking structure information, has higher repeated measure stability, or makes a greater contribution to model prediction. (c) For highly correlated pairs of clinical baseline features, retain at least one of the following factors: clinical importance, repeated measures stability, and model prediction contribution; (d) Features that are not relevant enough to be removed but have independent clinical significance shall be retained.

7. The method according to claim 1, characterized in that, The machine learning classification model with added regularization constraints includes at least one of the following: a logistic regression model with L2 regularization, a support vector machine model with radial basis function kernel function and L2 regularization, a random forest model with limited maximum tree depth, or a limit gradient boosting tree model with added L1 and L2 regularization terms; wherein, the classification label of the machine learning classification model is determined by a preset clinical reference standard, which includes the assessment conclusion of the Neonatal Oral Motor Assessment Scale (NOMAS) or the assessment conclusion of the Preterm Infant Oral Feeding Readiness Assessment Scale (PIOFRA).

8. The method according to claim 1, characterized in that, The method further includes: evaluating the machine learning classification model using hierarchical five-fold cross-validation, independently calculating and reporting the area under the ROC curve, accuracy, recall, specificity, and F1 score; and using the DeLong test to assess whether the differences in the area under the ROC curve among different models are statistically significant.

9. A neonatal oral motor function auxiliary assessment system, used to implement the neonatal oral motor function assessment method as described in any one of claims 1 to 8, characterized in that, The system includes: The data acquisition module is used to acquire raw data of pressure signals from non-nutritive sucking in newborns; The preprocessing module is used to perform preprocessing steps; The multi-strategy window capture module is used to execute window capture methods; The peak identification module is used to perform local extremum detection to identify effective positive pressure peaks and effective negative pressure peaks; The sucking burst determination and feature extraction module is used to identify the sucking burst structure and extract NNS dynamic features; The feature processing module is used to perform correlation analysis and redundancy elimination steps; The regularization evaluation modeling module is used to perform machine learning modeling, parameter tuning, probability calibration, and model evaluation. The evaluation result output module is used to provide auxiliary evaluation probabilities, risk levels, or signal quality prompts based on the model output.

10. An electronic device comprising a memory and a processor, characterized in that: The memory stores a computer program; when the processor reads and executes the computer program, it implements the neonatal oral motor function auxiliary assessment method according to any one of claims 1 to 8.

Images