Ovarian function prediction method and equipment based on multimodal data

Abstract

This invention relates to the field of physiological data processing technology, and discloses a method and device for predicting ovarian function based on multimodal data. The method includes: collecting a user's physiological time-series data through a wearable device, and obtaining standard scale scores obtained by the user through completing a standard questionnaire; performing anomaly detection on the physiological time-series data to trigger the collection of immediate questionnaire data, which reflects the user's instantaneous state when the physiological time-series data is abnormal; extracting features from the physiological time-series data to obtain physiological features, and extracting features from the standard scale scores and immediate questionnaire data to obtain subjective perception features; fusing the physiological features and subjective perception features to obtain a high-dimensional time-series feature vector set; and using a time series prediction model to predict ovarian function using the high-dimensional time-series feature vector set to obtain the ovarian function prediction result. This invention improves the accuracy and personalized adaptability of ovarian function prediction.

Description

Technical Field

[0001] This invention relates to the field of physiological data processing technology, and more specifically, to a method and device for predicting ovarian function based on multimodal data. Background Technology

[0002] Ovarian reserve (OR) refers to the number of primordial follicles remaining in the ovary and their ability to develop into mature oocytes. It is a core indicator for assessing a woman's fertility potential, reproductive health status, and predicting the timing of menopause, and has important clinical value for natural pregnancy, assisted reproductive decision-making, and health management throughout the entire life cycle.

[0003] Currently, the gold standard for clinical assessment of ovarian reserve mainly includes serum anti-Müllerian hormone (AMH) testing and antral follicle count (AFC) by ultrasound. However, these methods are all single-spot measurements and cannot capture the menstrual cycle fluctuations and diurnal rhythms of hormones; at the same time, they ignore the transient regulatory effects of individual differences and psychological factors on ovarian function, thus leading to inaccurate individualized predictions of ovarian function. Summary of the Invention

[0004] In view of this, the present invention provides a method and device for predicting ovarian function based on multimodal data, in order to solve the problem of inaccurate individualized prediction of ovarian function.

[0005] In a first aspect, the present invention provides a method for predicting ovarian function based on multimodal data. The method includes: collecting a user's physiological time-series data through a wearable device and obtaining standard scale scores obtained by the user through completing a standard scale; performing anomaly detection on the physiological time-series data to trigger the collection of immediate questionnaire data, which reflects the user's instantaneous state when the physiological time-series data is abnormal; extracting features from the physiological time-series data to obtain physiological features, and extracting features from the standard scale scores and immediate questionnaire data to obtain subjective perception features; fusing the physiological features and subjective perception features to obtain a high-dimensional time-series feature vector set; and using a time series prediction model to predict ovarian function using the high-dimensional time-series feature vector set to obtain the ovarian function prediction result.

[0006] In one optional implementation, anomaly detection is performed on physiological time-series data to trigger the collection of immediate questionnaire data, including: obtaining the user's dynamic personalized baseline; performing anomaly detection on the physiological time-series data based on the dynamic personalized baseline, the anomaly detection including a first anomaly detection condition and a second anomaly detection condition, the first anomaly detection condition being whether the proportion of each data point in the physiological time-series data deviating from the dynamic personalized baseline within a preset time exceeds a preset absolute threshold, and the second anomaly detection condition being whether the slope of the time-series trend of each data point in the physiological time-series data within a preset time window exceeds the historical normal fluctuation rate; if any data point in the physiological time-series data satisfies at least one of the first anomaly detection condition and the second anomaly detection condition, the collection of immediate questionnaire data is triggered.

[0007] In one optional implementation, the physiological time-series data includes resting heart rate, heart rate variability data, body temperature data, blood oxygen saturation data, body movement data, and sleep stage data. Feature extraction is performed on the physiological time-series data to obtain physiological features, including: feature extraction from resting heart rate data to obtain the nighttime mean resting heart rate and the daytime mean resting heart rate; feature extraction from heart rate variability data to obtain autonomic nervous system regulation features, including time-domain and frequency-domain features; feature extraction from body temperature data to obtain circadian rhythm features, including the diurnal amplitude, diurnal phase, and time of the nighttime minimum of body temperature; feature extraction from blood oxygen saturation data to obtain the nighttime minimum blood oxygen value and the number of blood oxygen drop events; feature extraction from body movement data to obtain daytime activity level and activity intensity; and feature extraction from sleep stage data to obtain deep sleep duration, REM sleep duration, light sleep duration, and the number of nighttime awakenings.

[0008] In one optional implementation, feature extraction is performed on standard scale scoring data and immediate questionnaire data to obtain subjective perception features, including: feature extraction of standard scale scoring data to obtain long-term statistical features, which include the average value and volatility of standard scale scoring data over a preset long period; feature extraction of immediate questionnaire data to obtain short-term dynamic features, which include the average value and volatility of immediate questionnaire scores over a preset short period, and the cross-correlation coefficient between immediate questionnaire scores and the intensity of physiological events; and combining the long-term statistical features and short-term dynamic features to obtain subjective perception features.

[0009] In one optional implementation, ovarian function is predicted using a time series prediction model on a high-dimensional time series feature vector set to obtain ovarian function prediction results. This includes: adding positional encoding to the high-dimensional time series feature vector set using the time series prediction model to obtain a positionally encoded time series feature sequence; using the self-attention mechanism of the time series prediction model to process the positionally encoded time series feature sequence in parallel, capturing long-term dependencies between different time steps and cross-modal associations between physiological features and subjectively perceived features to obtain a deep time series feature representation; and using the time series prediction model to perform a linear mapping on the deep time series feature representation to obtain predicted values ​​of ovarian reserve markers, which serve as the ovarian function prediction results.

[0010] In one optional implementation, the training process of the time series prediction model includes: acquiring a first training dataset, which includes physiological time-series training data of sample users, standard scale rating training data, instant questionnaire training data, and gold standard label data; aligning the multimodal data in the first training dataset according to time windows to obtain an aligned multimodal time-series feature vector set; adding positional encoding to the multimodal time-series feature vector set, and training the initial time series prediction model through parallel processing and linear mapping using a self-attention mechanism to obtain a first predicted value; calculating the root mean square error (RMSE) of the first predicted value and the gold standard label data to obtain a first RMSE result; and optimizing the model parameters through backpropagation based on the first RMSE result until the model converges to obtain the time series prediction model.

[0011] In one optional implementation, the method of using a time series prediction model to predict ovarian function from a high-dimensional time series feature vector set to obtain ovarian function prediction results further includes: using the time series prediction model to tile the high-dimensional time series feature vector set to obtain matrix-format features; using the time series prediction model to perform feature selection on the matrix-format features using a multi-objective evolutionary feature selection method to obtain first-selected features; using the time series prediction model to perform feature selection on the matrix-format features using Fourier series transform to obtain second-selected features; using the time series prediction model to concatenate the first-selected features and the second-selected features, and introducing multi-order lag variables to incorporate physiological features and subjective perception features within the historical time window into the current moment to obtain an extended feature set; using the time series prediction model to evaluate the predictive performance of the extended feature set on multiple prediction tasks through statistical tests and multiple cross-validation; using the time series prediction model to output predicted values ​​of ovarian reserve markers based on the extended feature set through an autoregressive method, and using multi-criteria decision-making to select the predicted values ​​of ovarian reserve markers to obtain ovarian function prediction results.

[0012] In one optional implementation, the training process of the time series prediction model further includes: acquiring a second training dataset, which includes physiological time-series training data of sample users, standard scale rating training data, instant questionnaire training data, and gold standard label data; performing feature tiling on the multimodal data in the second training dataset to obtain matrix-formatted training features; using a multi-objective evolutionary feature selection method to filter the matrix-formatted training features to obtain a first candidate feature subset; using Fourier series transform to filter the matrix-formatted training features to obtain a second candidate feature subset; concatenating the first candidate feature subset with the second candidate feature subset and introducing multi-order lagged variables to construct an extended training feature set; evaluating the predictive performance of the extended training feature set using multiple cross-validation and statistical tests; determining the optimal extended feature set configuration based on the cross-validation results through multi-criteria decision-making; and training the model using a regression method based on the optimal extended feature set configuration until the model converges to obtain the time series prediction model.

[0013] The present invention provides a method for predicting ovarian function based on multimodal data, which further includes: using a time series prediction model to perform risk analysis on the ovarian function prediction results to obtain risk classification results and fertility index. The risk classification results indicate the risk level of the user's ovarian function, and the fertility index is used to provide the user's fertility score.

[0014] Secondly, the present invention provides an ovarian function prediction device based on multimodal data, comprising: a processor and a memory connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to cause the processor to perform any of the above-mentioned ovarian function prediction methods based on multimodal data.

[0015] This invention provides a method for predicting ovarian function based on multimodal data. First, it continuously collects high-precision physiological time-series data using wearable devices, supplemented by standard scale scoring data, achieving long-term, non-invasive, and personalized data collection at home. Next, it triggers an immediate questionnaire using real-time anomaly detection, introducing instantaneous ecological assessment into ovarian function prediction for the first time. This establishes a real-time correlation between physiological abnormalities and the user's immediate psychological experience, improving the accuracy of ovarian function prediction. Then, through multimodal feature extraction and fusion, objective physiological indicators and subjective psychological experiences are transformed into structured representations, fully utilizing the complementarity of these two types of information to provide more comprehensive feature input for ovarian function prediction. Finally, it uses a time-series prediction model to learn long-term temporal dependencies and cross-modal associations, enabling early identification of the risk of ovarian function decline before substantial abnormalities occur in clinical indicators, achieving proactive early warning of changes in ovarian function. This method balances low-cost, home-based compliance with high-precision individualized prediction, improving the accuracy and personalized adaptability of ovarian function prediction. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating a method for predicting ovarian function based on multimodal data according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating another method for predicting ovarian function based on multimodal data according to an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] According to an embodiment of the present invention, an embodiment of a method for predicting ovarian function based on multimodal data is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0020] This embodiment provides a method for predicting ovarian function based on multimodal data, which can be used on the aforementioned mobile terminals, such as mobile phones and tablets. Figure 1 This is a flowchart of an ovarian function prediction method based on multimodal data according to an embodiment of the present invention, such as... Figure 1 As shown, the process includes the following steps: Step S101: Collect the user's physiological time-series data through the wearable device and obtain the standard scale score data obtained by the user by filling in the standard scale.

[0021] In this embodiment, physiological time-series data is continuously and non-invasively collected through wearable devices (such as smart rings) in the user's natural daily life. The smart ring can be worn at the base of the finger, and its photoplethysmography (PPG) signal-to-noise ratio is significantly higher than that of wrist-worn devices, achieving medical-grade pulse wave extraction accuracy. Continuous skin temperature measurement at the fingertips has a resolution of 0.01℃, capturing minute temperature changes mediated by the hypothalamus-pituitary-ovarian axis. Simultaneously, the compact structure of the smart ring effectively reduces motion artifacts, ensuring the integrity of long-term data. Furthermore, the smart ring is lightweight and discreet, allowing users to monitor at home for extended periods, avoiding frequent medical visits. In addition, physiological time-series data can be collected using purely time-triggered methods (e.g., three times a day, morning, noon, and evening). This method is simple to implement and has high data processing efficiency. Wearable devices can also utilize built-in sensors in mobile phones or fitness trackers to reduce user costs and facilitate wider adoption.

[0022] In this embodiment, the standard scale scoring data is obtained through questionnaires pushed via a mobile application linked to the wearable device. The questionnaires include the Pittsburgh Sleep Quality Index (PSQI) and the Symptom Checklist-90, an internationally recognized mental health assessment tool. Users complete the questionnaires and receive corresponding total and factor scores, i.e., standard scale scoring data, which reflects the baseline of sleep quality and mental health at different stages. This embodiment collects users' objective physiological and subjective psychological data, which can be used for subsequent multimodal fusion and accurate prediction of ovarian function. It should be noted that standard scale scoring data is only collected when pushed to the mobile application at a preset frequency (e.g., weekly or monthly); no data is available outside of push notification times. If the user does not fill in the questionnaire during the push notification time, the data for that time will also be empty. Null values ​​are not included in the statistical calculations during subsequent feature extraction.

[0023] Step S102: Perform anomaly detection on physiological time series data to trigger the collection of real-time questionnaire data. The real-time questionnaire data is used to reflect the instantaneous state of the user when the physiological time series data is abnormal.

[0024] This embodiment requires real-time anomaly analysis of continuously collected physiological time-series data. When any indicator in the physiological time-series data shows a significant deviation from the user's normal state (e.g., a sudden and drastic change in value or a deviation from the baseline for several consecutive days), a short, instant questionnaire is immediately pushed to the mobile application. This questionnaire contains only a few quick-response questions to collect the user's subjective feelings at the moment the physiological abnormality occurred. This data is called instant questionnaire data. It is also subjective rating data, but unlike standard scales, it is event-driven, high-frequency, and instantaneous, accurately reflecting the dynamic relationship between specific physiological stress events and psychological states. By capturing this coupling relationship between physiological abnormalities and immediate psychological feedback, the instantaneous impact of emotions and stress on ovarian function can be further explored, making up for the blind spots of traditional single physiological indicators or retrospective questionnaires.

[0025] Step S103: Extract features from physiological time-series data to obtain physiological features, and extract features from standard scale scoring data and real-time questionnaire data to obtain subjective perception features.

[0026] Before feature extraction, this embodiment can preprocess the physiological time-series data, including denoising, outlier removal, and motion artifact repair, to improve signal quality. Based on this, quantitative indicators characterizing the user's physiological state are extracted from the cleaned physiological data as physiological features. Simultaneously, quantitative indicators reflecting the user's stage-specific psychological state and instantaneous subjective experience are extracted from standard scale rating data (regularly completed) and real-time questionnaire data (event-triggered), respectively, jointly constituting subjective perception features. The purpose of this step is to transform the raw multimodal data into feature representations with clear biological or psychological significance, remove redundant noise, and provide structured, low-redundancy input for subsequent multimodal fusion and ovarian function prediction.

[0027] Step S104: The physiological characteristics and subjective perception characteristics are fused to obtain a high-dimensional temporal feature vector set.

[0028] This embodiment combines and aligns physiological characteristics and subjective perception characteristics according to a unified time benchmark, forming a high-dimensional temporal feature vector set containing multi-dimensional information. This fusion process aims to integrate objective physiological indicators and subjective psychological assessment data into a joint representation, enabling subsequent models to simultaneously utilize the complementary information of the two types of features to more comprehensively depict the dynamic relationship between the user's physiological and psychological state and ovarian function.

[0029] Step S105: Use a time series prediction model to predict ovarian function from a high-dimensional time series feature vector set, and obtain the ovarian function prediction results.

[0030] This embodiment inputs a high-dimensional time-series feature vector set into a pre-trained time series prediction model. The model automatically learns the temporal dependencies within the feature vectors and the cross-modal correlations between physiological features and subjectively perceived features, ultimately outputting a prediction result related to ovarian function. This prediction result can be used to assess a user's current ovarian status or future trends.

[0031] This embodiment provides a multimodal data-based method for predicting ovarian function. First, it continuously collects high-precision physiological time-series data using wearable devices, supplemented by standard scale scoring data, achieving long-term, non-invasive, and personalized data collection at home. Next, it triggers an immediate questionnaire using real-time anomaly detection, introducing instantaneous ecological assessment into ovarian function prediction for the first time. This establishes a real-time correlation between physiological abnormalities and the user's immediate psychological experience, improving the accuracy of ovarian function prediction. Then, through multimodal feature extraction and fusion, objective physiological indicators and subjective psychological experiences are transformed into structured representations, fully utilizing the complementarity of these two types of information to provide more comprehensive feature input for ovarian function prediction. Finally, it uses a time-series prediction model to learn long-term temporal dependencies and cross-modal associations, enabling early identification of ovarian function decline risks before substantial abnormalities occur in clinical indicators, achieving proactive early warning of changes in ovarian function. This method balances low-cost home-based adherence with high-precision individualized prediction, improving the accuracy and personalized adaptability of ovarian function prediction.

[0032] In some alternative implementations, the physiological time-series data in step S101 includes resting heart rate, heart rate variability data, body temperature data, blood oxygen saturation data, body movement data, and sleep stage data.

[0033] Among them, resting heart rate reflects the baseline cardiac activity level; heart rate variability data characterizes the balance of the autonomic nervous system; body temperature data is used to monitor diurnal rhythms and basal body temperature changes; blood oxygen saturation data indicates the status of respiratory and circulatory function; body movement data is used to quantify the intensity and pattern of physical activity; and sleep stage data distinguishes different sleep stages such as light sleep, deep sleep, and REM sleep, and is used to assess sleep structure and quality. These indicators together constitute a multidimensional basis for assessing the user's physiological status, providing comprehensive and objective data support for subsequent ovarian function prediction.

[0034] In some optional implementations, step S102 involves anomaly detection of physiological time-series data to trigger the collection of immediate questionnaire data, specifically including: Step S1021: Obtain the user's dynamic personalization baseline.

[0035] This embodiment establishes a dynamic, personalized baseline for users based on long-term, continuously collected multidimensional physiological time-series data, using multidimensional physiological indicators (including autonomic nervous system regulation, circadian rhythm, sleep structure and quality, and exercise and metabolism indicators). This baseline reflects the normal fluctuation range of users at different time scales (e.g., circadian rhythm, menstrual cycle). By constructing a personalized baseline, this embodiment avoids misjudgments caused by using fixed thresholds for the general population, providing an individualized reference benchmark for subsequent anomaly detection, thereby significantly improving the targeting and accuracy of anomaly detection.

[0036] Step S1022: Anomaly detection is performed on the physiological time series data based on the dynamic personalized baseline. The anomaly detection includes a first anomaly detection condition and a second anomaly detection condition. The first anomaly detection condition is to detect whether the proportion of each data in the physiological time series data deviating from the dynamic personalized baseline within a preset time exceeds a preset absolute threshold. The second anomaly detection condition is to detect whether the temporal trend slope of each data in the physiological time series data within a preset time window exceeds the historical normal fluctuation rate.

[0037] This embodiment performs real-time anomaly analysis on continuously collected physiological time-series data. Based on the user's dynamic personalized baseline, and through two rules—absolute threshold triggering and time-series trend deterioration triggering—a short questionnaire is immediately triggered when any abnormality is detected, to collect the user's current instantaneous subjective feelings.

[0038] Specifically, the physiological time-series data in this embodiment covers four types of indicators, and the abnormality determination method for each type of indicator is as follows: Autonomic nervous system regulation indicators include Heart Rate Variability (HRV) related indicators (Root Mean Square of Successive Differences (RMSSD), Standard Deviation of Normal-to-Normal Intervals (SDNN), and Low Frequency / High Frequency Ratio (LF / HF) in the time domain) and Resting Heart Rate (RHR). For HRV indicators, the first anomaly detection condition is that RMSSD or SDNN is below 25% of the individual mean for three consecutive days, or the LF / HF ratio is above 50% of the individual mean for three consecutive days. The second anomaly detection condition is a continuously decreasing slope of RMSSD or SDNN, or a continuously increasing slope of LF / HF, calculated using a sliding window. If the absolute value of the slope exceeds the historical normal volatility (e.g., the 95% confidence interval of the slope over the past 90 days), it is triggered. For resting heart rate, the first abnormality detection condition is that the nighttime resting heart rate (the average heart rate during the nighttime sleep period) has increased by more than 5 beats / min relative to the individual baseline for 3 consecutive days; the second abnormality detection condition is that the resting heart rate has shown an upward trend for 7 consecutive days and the slope exceeds the historical normal fluctuation rate.

[0039] Biological clock and circadian rhythm indicators: This category includes baseline fluctuations in core body temperature / skin temperature (BBT / Skin Temperature) and diurnal temperature amplitude and phase lag. For BBT / Skin Temperature baseline fluctuations, the first abnormal detection criterion is a deviation of the nighttime mean skin temperature from the estimated baseline for the current menstrual cycle for several consecutive days (e.g., insufficient post-ovulation temperature rise or abnormal duration of low temperature during the luteal phase). For diurnal temperature amplitude and phase lag, the second abnormal detection criterion is a narrowing of the diurnal temperature range or an abnormally delayed trough.

[0040] Sleep structure and quality indicators: This category includes nighttime wake-up frequency (WASO) and duration, as well as sleep structure imbalance (the ratio of deep sleep to rapid eye movement (REM) sleep). For nighttime wake-up frequency, the first abnormality detection criterion is that the number of wake-ups significantly exceeds the individual's historical average for several consecutive days, or the duration of a single wake-up is excessively long; the second abnormality detection criterion is that the wake-up frequency shows a continuous increasing trend and the rate of deterioration exceeds the historical normal fluctuation rate. For sleep structure imbalance, the first abnormality detection criterion is that the proportion of deep sleep or REM sleep is significantly lower than the individual's historical baseline for several consecutive days; the second abnormality detection criterion is that the proportion of deep sleep or REM sleep shows a continuous decreasing trend and the rate of deterioration exceeds the historical normal fluctuation rate.

[0041] Exercise and metabolic indicators: These indicators are daily activity levels and energy expenditure (calculated using a triaxial accelerometer to measure daily steps, activity intensity, and metabolic equivalent). The first abnormality detection condition is that daily activity levels are significantly lower than the individual's baseline for several consecutive days; the second abnormality detection condition is that activity levels show a downward trend for several consecutive days and the rate of deterioration exceeds the historical normal fluctuation rate.

[0042] Step S1023: If any data in the physiological time series data satisfies at least one of the first anomaly detection condition and the second anomaly detection condition, then the collection of real-time questionnaire data is triggered.

[0043] In this embodiment, when the first or second anomaly detection condition corresponding to any indicator in step S1022 is met, a short questionnaire will be pushed through the mobile application immediately or within a preset golden time window (e.g., within 5 minutes after the anomaly occurs) to collect instantaneous subjective data such as the user's current psychological stress, emotional fluctuations, fatigue, pain symptoms, and libido, ensuring that the subjective feelings and objective physiological abnormal events are accurately aligned in time.

[0044] This embodiment first establishes a dynamic, personalized baseline, using the normal fluctuation range of the user's own multidimensional physiological indicators as a reference benchmark. This avoids misjudgments caused by fixed thresholds for the general population, significantly improving the individualized adaptation and accuracy of anomaly detection. Next, it employs two complementary rules: a first anomaly detection condition and a second anomaly detection condition. The former can capture acute stress events such as a sudden drop in heart rate variability and a surge in awakenings, while the latter can identify progressive decline trends such as a sustained increase in heart rate, disordered body temperature rhythm, and deterioration of sleep structure. The combination of these two rules comprehensively covers physiological anomaly patterns at different time scales, greatly enhancing the sensitivity and early warning capabilities of anomaly detection. When any detection condition is met, an immediate questionnaire is precisely triggered, ensuring that subjective feelings and objective physiological anomalies are accurately aligned in time, avoiding the user burden and data redundancy caused by blind push notifications. Overall, this embodiment achieves a personalized, highly sensitive, and low-redundancy anomaly triggering mechanism, providing a high-quality event-level data foundation for subsequent dynamic coupling modeling of physiological and psychological factors and ovarian function prediction, thereby effectively improving the accuracy of ovarian function prediction.

[0045] In some optional implementations, the process of extracting features from physiological time-series data in step S103 to obtain physiological features specifically includes: Step S103a1: Extract features from the resting heart rate data to obtain the mean resting heart rate at night and the mean resting heart rate during the day.

[0046] This embodiment extracts the mean heart rate during nighttime (e.g., during sleep) and daytime (during waking and activity) periods from continuously acquired resting heart rate sequences. The mean resting heart rate at night reflects baseline autonomic nervous system tone, while the mean resting heart rate during daytime reflects the cardiac workload level during daily activities. By distinguishing between daytime and nighttime resting heart rates, the diurnal differences in autonomic nervous system function and their long-term trends can be captured more precisely, providing a basic indicator for subsequent assessment of the impact of stress on ovarian function.

[0047] Step S103a2: Feature extraction is performed on the heart rate variability data to obtain autonomic nervous system regulation features, which include time domain features and frequency domain features.

[0048] This embodiment extracts time-domain features (e.g., root mean square difference RMSSD) and frequency-domain features (e.g., low-frequency / high-frequency ratio LF / HF) from heart rate variability data. Time-domain features reflect parasympathetic activity, while frequency-domain features reflect the sympathetic-parasympathetic balance. These features can quantify the impact of psychological stress and emotional fluctuations on the autonomic nervous system, thereby revealing the link between autonomic nervous regulation and reproductive endocrine function, and providing neurophysiological markers for predicting ovarian dysfunction.

[0049] Step S103a3: Extract features from body temperature data to obtain circadian rhythm features, which include the diurnal amplitude, diurnal phase, and time of the lowest point at night.

[0050] This embodiment extracts three core parameters of circadian rhythm from continuous body temperature data: amplitude (diurnal temperature range), phase (the time of peak or trough), and the time of the nighttime trough. These parameters reflect the stability of the biological clock and the rhythmic regulation capacity of the hypothalamic-pituitary-ovarian axis (HPO). Abnormalities in rhythmic characteristics (e.g., decreased amplitude, phase delay) can provide early indications of endocrine rhythm disorders, offering a chronobiological basis for proactive early warning of ovarian dysfunction.

[0051] Step S103a4: Feature extraction is performed on the blood oxygen saturation data to obtain the lowest blood oxygen value at night and the number of blood oxygen drop events.

[0052] This embodiment extracts the lowest blood oxygen value and blood oxygenation drop events (e.g., the number of times blood oxygen levels drop) from the nocturnal blood oxygen saturation sequence. These indicators are used to assess sleep apnea and the degree of nocturnal hypoxia. Long-term nocturnal hypoxia may affect ovarian reserve through oxidative stress and inflammatory responses; therefore, this feature can serve as an auxiliary physiological indicator for assessing ovarian function.

[0053] Step S103a5: Extract features from the body movement data to obtain the daytime activity level and intensity.

[0054] This embodiment extracts total daily activity levels (e.g., steps, activity counts) and activity intensity (e.g., activity duration) from triaxial accelerometer data. Activity levels and intensity reflect a user's daily physical exertion levels; low activity levels or decreased intensity may indicate physical fatigue or chronic stress. This feature helps to analyze the impact of lifestyle on ovarian function and serves as a basis for personalized health interventions.

[0055] Step S103a6: Extract features from sleep stage data to obtain deep sleep duration, REM sleep duration, light sleep duration, and number of nighttime awakenings.

[0056] This embodiment uses a sleep staging algorithm to divide the entire night's sleep into deep sleep, rapid eye movement (REM) sleep, light sleep, and wakefulness stages, and extracts the duration and number of awakenings for each stage. Deep sleep and REM sleep are crucial for hormone secretion (such as growth hormone and melatonin) and neural repair, and their structural abnormalities (such as shortened deep sleep and REM sleep ratio imbalance) are closely related to HPO axis dysfunction. Sleep structure characteristics provide important information at the central nervous system level for assessing ovarian function.

[0057] This embodiment first achieves a refined, hierarchical assessment of autonomic nervous system function by extracting the mean resting heart rate between day and night, enhancing the ability to characterize basic physiological states. Secondly, it extracts time-domain and frequency-domain autonomic nervous system regulatory features from heart rate variability data, quantifying the neural pathways by which psychological stress affects reproductive endocrine function. Next, it establishes a quantitative bridge between circadian rhythm characteristics of body temperature and the regulation of the hypothalamus-pituitary-ovarian axis. Then, it supplements the indirect effects of oxidative stress and lifestyle on ovarian function using nighttime blood oxygenation characteristics and daytime activity levels. Finally, it reveals the connection between central nervous system sleep regulation and the reproductive axis through sleep structure characteristics. The extraction of these multi-level physiological features transforms raw signals into quantitative indicators with clear biological significance, providing structured, low-redundancy, and multi-source data input for subsequent multimodal fusion and ovarian function prediction, significantly improving the accuracy of ovarian function prediction.

[0058] In some optional implementations, step S103, which involves extracting features from the standard scale scoring data and the immediate questionnaire data to obtain subjective perception features, specifically includes: Step S103b1: Extract features from the standard scale scoring data to obtain long-term statistical features, which include the average value and volatility of the standard scale scoring data over a preset long period.

[0059] This embodiment aggregates user-reported scores from standard scales (e.g., the Pittsburgh Sleep Quality Index (PSQI) and the Symptom Checklist-90) over a preset long period (e.g., one month), calculating the arithmetic mean and mean squared deviation (volatility) of all scores within that period. The mean reflects the user's overall sleep quality or mental health level within that period, while the volatility characterizes its stability across periods. This long-term statistical feature can capture slowly changing psychological states such as chronic stress, persistent low mood, or long-term sleep disorders, providing a stable subjective psychological baseline for ovarian function prediction and avoiding interference from momentary emotional fluctuations on long-term trends.

[0060] Step S103b2: Extract features from the real-time questionnaire data to obtain short-term dynamic features. The short-term dynamic features include the average value and volatility of the real-time questionnaire scores within a preset short period, as well as the cross-correlation coefficient between the real-time questionnaire scores and the intensity of physiological events.

[0061] This embodiment statistically analyzes instantaneous questionnaire scores triggered by events (such as psychological stress, mood swings, fatigue, pain symptoms, and libido) over a preset short period (e.g., one week). The average and mean squared deviation (volatility) of all instantaneous scores within this period are calculated, reflecting the recent average subjective experience and the intensity of its fluctuations. Simultaneously, the Pearson correlation coefficient between the event intensity (e.g., the ratio of the heart rate spike to baseline) of each physiological event (e.g., a spike in heart rate, a sudden drop in HRV, or an abnormal increase in awakenings) and the corresponding instantaneous questionnaire score is calculated to quantify the synchronous coupling strength between the user's subjective feelings and objective physiological abnormalities. These short-term dynamic characteristics can capture real-time changes in acute stress responses, emotional instability, and the physiological-psychological relationship, providing crucial data for revealing the transient impact of psychological factors on ovarian function.

[0062] Step S103b3: Combine long-term statistical features and short-term dynamic features to obtain subjective perception features.

[0063] This embodiment vertically concatenates or weights short-term dynamic features to form a unified subjective perception feature vector. This combined feature vector includes both chronic psychological baseline information (such as sleep quality and emotional stability over weeks to months) and instantaneous stress response patterns (such as event-driven stress responses and their coupling with physiological abnormalities), comprehensively depicting the user's multi-scale subjective state from macro to micro levels. This multi-scale fusion improves the subsequent model's ability to identify psychological and physiological coupling patterns, thereby enhancing the individualized adaptability of ovarian function prediction.

[0064] This embodiment first establishes a stable and reliable baseline for mental health and sleep quality by calculating the long-term average and volatility of standard scale scores. This provides a quantitative basis for capturing the cumulative impact of chronic psychological stress on ovarian function, avoiding interference from short-term, accidental factors. Secondly, by calculating the short-term average and volatility of immediate questionnaire scores, it reflects the user's recent emotional instability and stress fluctuations in real time, enabling sensitive identification of acute psychological stress events. Next, by introducing the cross-correlation coefficient between immediate questionnaire scores and the intensity of physiological events, it accurately characterizes the user's personalized psychological response patterns when facing physiological abnormalities. Finally, by combining long-term statistical characteristics with short-term dynamic characteristics, it achieves complementary fusion of subjective information across multiple time scales, preserving both macro-trends and capturing micro-dynamics.

[0065] In some alternative implementations, step S103 may extract only simple statistical features (e.g., daily average heart rate, weekly average body temperature, etc.) from the physiological time-series data. In step S104, these statistical features are concatenated to form a low-dimensional feature vector. Then, a conventional linear model (e.g., linear regression, logistic regression) is used to predict ovarian function. The advantages of this alternative are minimal computational overhead, fast model training and inference speed, and the ability to be directly deployed on edge devices (e.g., smart rings or mobile phones) without relying on cloud computing power, enabling real-time or near real-time local prediction.

[0066] In some optional implementations, step S105, which uses a time series prediction model to predict ovarian function from a high-dimensional time-series feature vector set to obtain the ovarian function prediction result, specifically includes: Step S105a1: Use a time series prediction model to add positional encoding to the high-dimensional time series feature vector set to obtain a time series feature sequence with positional encoding.

[0067] In this embodiment, the time series prediction model adopts a Transformer-based architecture. First, positional encoding is added to the high-dimensional temporal feature vector set (containing multi-dimensional time series of physiological and subjectively perceived features) generated in step S104. Since the Transformer's self-attention mechanism itself lacks temporal order perception capability, positional encoding uses sine / cosine functions or learnable embedding vectors to assign unique sequence position information to each time step, thereby preserving the temporal order relationship in the original data. This step ensures that the model can correctly distinguish input features at different times, laying the foundation for subsequently capturing long-term dependencies across time steps.

[0068] Step S105a2: The self-attention mechanism of the time series prediction model is used to process the temporal feature sequence with position encoding in parallel, capture the long-term dependency between different time steps and the cross-modal association between physiological features and subjective perception features, and obtain a deep temporal feature representation.

[0069] This embodiment inputs a temporal feature sequence with positional encoding into a Transformer encoder and utilizes a multi-head self-attention mechanism to compute the attention weights between any two time steps in the sequence in parallel. This mechanism can directly model long-range dependencies across menstrual cycles and circadian rhythms (e.g., the impact of physiological fluctuations from weeks ago on current ovarian status), while automatically uncovering cross-modal implicit associations between physiological features (e.g., HRV, body temperature) and subjectively perceived features (e.g., stress scores, sleep quality). Through the stacking of multiple encoding layers, the model progressively extracts abstract deep temporal feature representations, embedding complex nonlinear dynamic patterns between the endocrine system and psychological stress under HPO axis regulation. Parallel computation of self-attention significantly improves training efficiency and avoids the gradient vanishing problem in long sequences.

[0070] Step S105a3: The time series prediction model is used to perform a linear mapping on the deep time series feature representation to obtain the predicted values ​​of ovarian reserve markers, which are used as the prediction results of ovarian function.

[0071] This embodiment inputs the deep temporal feature representation (i.e., global encoding of multimodal temporal data) output by the Transformer encoder into the decoder. The decoder uses an autoregressive approach to generate a prediction sequence step-by-step: in each decoding step, the decoder combines the encoder output with previously generated historical predictions, and through masked self-attention and encoder-decoder attention modules, progressively outputs the predicted values ​​of ovarian reserve markers for the next time step (e.g., predicted values ​​of anti-Müllerian hormone (AMH) concentration, antral follicle count (AFC), or follicle-stimulating hormone (FSH) level). Finally, the decoder generates a complete prediction sequence, with the output of the last time step (or the output at a specified time point) serving as the ovarian function prediction result. This decoding process can fully utilize the long-term dependency and cross-modal association information extracted by the encoder, making it suitable for prospective clinical analysis.

[0072] This embodiment first preserves the temporal order information of the multimodal temporal feature vector set through positional encoding, providing a sequence structure foundation for the self-attention mechanism. Secondly, leveraging the parallel processing capability of the self-attention mechanism, it directly captures the physiological and psychological dynamic dependencies across long time steps. Next, through nonlinear transformation of a multi-layer encoder, it deeply mines the hidden coupling patterns between HRV, body temperature, sleep structure, and EMA subjective scores, achieving high-order modeling of the complex regulatory laws of the hypothalamus-pituitary-ovarian axis. Finally, by outputting predicted values ​​of ovarian reserve markers through linear mapping, it completes high-precision regression from multimodal data to clinical indicators. This improves the accuracy and response speed of long-term temporal prediction, making it particularly suitable for ovarian function assessment tasks that require capturing physiological and psychological evolution over several months or even across menstrual cycles.

[0073] In some alternative implementations, the Transformer backbone network in step S105a1 can be replaced with a Long Short-Term Memory (LSTM) network, a Temporal Convolutional Network (TCN), or a novel state-space model (such as Mamba). These network models have low computational resource consumption and a small number of parameters, and can effectively avoid small-sample overfitting when performing personalized federated learning fine-tuning on the edge, making them suitable for deployment on smart rings or mobile devices with limited computing power.

[0074] In some alternative implementations, the training process for the time series prediction model includes: Step 1: Obtain the first training dataset, which includes physiological time-series training data of sample users, standard scale rating training data, real-time questionnaire training data, and gold standard label data.

[0075] This embodiment collects a large amount of historical data from users to form a training set. The training data includes physiological time-series data (HRV, body temperature, sleep, etc. continuously collected by wearable devices), standard scale scoring training data (periodic scores such as PSQI and SCL-90), and instantaneous questionnaire training data (instantaneous subjective scores triggered by events), all of which are high-frequency dynamic streams. The gold standard label data (AMH, AFC, FSH values) are sparse clinical indicators obtained from users' regular blood tests or ultrasound examinations at hospitals. Incorporating these data into the training set simultaneously provides a data foundation for subsequently establishing the mapping relationship between high-frequency features and low-frequency macro-labels.

[0076] Step 2: Align the multimodal data in the first training dataset according to the time window to obtain the aligned multimodal temporal feature vector set.

[0077] This embodiment uses the detection date of each gold standard label as an anchor point (e.g., a user's AMH test on April 1st), retrospectively and subsequently extending a reasonable clinical window period (e.g., 60 days prior to and 60 days after), labeling all high-frequency physiological time-series data and EMA data within this window with the same gold standard value. A large number of training samples are generated using a sliding window approach. Each sample contains a multimodal time-series feature vector within a continuous time window and corresponds to an ovarian reserve biomarker label. This method considers the relative stability of ovarian function (no drastic changes within six months), allowing sparse clinical ground truth values ​​to be allocated to high-frequency data within the window, effectively solving the time matching problem for data with different sampling frequencies.

[0078] Step 3: Add positional encoding to the multimodal time series feature vector set, and train the initial time series prediction model through parallel processing and linear mapping using a self-attention mechanism to obtain the first predicted value.

[0079] In this embodiment, the aligned multimodal temporal feature vector set is input into the initial Transformer model. First, positional encoding is added to each time step to preserve the temporal order. Then, the sequence within the entire window is processed in parallel through the encoder's self-attention mechanism to capture long-term dependencies across time steps (e.g., the impact of physiological fluctuations from several weeks ago on the current ovarian state) and cross-modal associations (e.g., the coupling between HRV decline and stress score). Finally, the first predicted value of the ovarian reserve marker corresponding to that window (e.g., AMH = 1.2 ng / mL) is obtained through the output layer (linear mapping or decoder). This step achieves an end-to-end mapping from high-frequency dynamic features to macroscopic clinical indicators.

[0080] Step four: Calculate the root mean square error (RMSE) of the first predicted value and the gold standard label data to obtain the first RMSE result.

[0081] This embodiment compares the predicted values ​​output by the model with the gold standard label corresponding to the window (e.g., the true AMH value), using either the root mean square error (RMSE) or the mean absolute error (MAE) as the loss function to calculate the difference between the two. RMSE is sensitive to large errors and can effectively drive the model to optimize prediction accuracy. This calculation process is repeated in the training batch, quantifying the prediction bias of the current model parameters.

[0082] Step 5: Based on the first root mean square error result, optimize the model parameters through backpropagation until the model converges, and obtain the time series prediction model.

[0083] This embodiment uses the backpropagation algorithm to calculate the gradient of the loss function with respect to the parameters of each layer of the model, and updates the weights of the Transformer (including the self-attention matrix, feedforward network parameters, positional encoding, etc.) through an optimizer (e.g., Adam). After multiple rounds of iterative training, the model converges when the RMSE on the validation set no longer decreases or reaches a preset number of rounds, finally yielding a time series prediction model that can be used for ovarian function prediction.

[0084] This embodiment first constructs a training set containing high-frequency multimodal data and sparse gold standard labels, providing comprehensive data support for learning the complex mapping between physiological and psychological dynamics and ovarian reserve. Secondly, a window-period resampling alignment strategy is employed to solve the time matching problem of data with different sampling frequencies, ensuring that high-frequency features are correctly associated with macro-labels and avoiding model learning bias caused by time misalignment. Next, the self-attention mechanism of the Transformer is used to process all time-series data within the window in parallel, fully capturing long-term dependencies across menstrual cycles and cross-modal coupling between physiological and psychological aspects, improving the model's ability to express nonlinear dynamic patterns. Then, the root mean square error loss function is used to accurately quantify prediction bias, driving the model towards higher accuracy. Finally, backpropagation and end-to-end training enable the model to automatically learn the optimal mapping function from multimodal time-series features to ovarian reserve markers. The entire training process ensures that the model can accurately predict ovarian function status based on users' daily wearable data and EMA data in practical applications, laying the algorithmic foundation for non-invasive, continuous, and personalized ovarian reserve assessment.

[0085] In some optional implementations, step S105, which uses a time series prediction model to predict ovarian function from a high-dimensional time-series feature vector set to obtain the ovarian function prediction result, further includes: Step S105b1: Use a time series prediction model to tile the high-dimensional time series feature vector set to obtain matrix-format features.

[0086] This embodiment performs a feature tiling operation on the high-dimensional temporal feature vector set generated in step S104, that is, expands or flattens it along the time axis, converting it into a two-dimensional matrix format (number of samples × number of features) that can be processed by conventional machine learning algorithms. This operation eliminates the rigid dependence of the input data on the original timestamp order, so that subsequent algorithms do not need to consider the sequence structure of time steps, reducing the requirements of the model for complex temporal modeling. At the same time, the tiling matrix format is compatible with traditional regression models (such as random forests, support vector machines, etc.), providing standardized input for subsequent feature engineering and model training.

[0087] Step S105b2: Using a time series prediction model, a multi-objective evolutionary feature selection method is employed to filter the matrix-format features and obtain the first selected features.

[0088] This embodiment employs a multi-objective evolutionary feature selection method to filter and reduce the dimensionality of the tiled matrix-formatted features. This method encodes each candidate feature subset as an individual, using maximizing prediction accuracy, minimizing the number of features, and minimizing computational cost as multiple conflicting optimization objectives. Through iterative evolutionary search (including genetic operations such as selection, crossover, and mutation), a set of optimal feature subset candidates is generated, from which the optimal feature subset is selected as the first screening feature. This selected feature provides key temporal dimensions for subsequent merging.

[0089] Step S105b3: Using the time series prediction model, Fourier series transform is used to filter the matrix format features to obtain the second filtered features.

[0090] This embodiment is another processing method parallel to step S105b2. This embodiment uses an improved Fourier series to perform frequency domain transformation on the matrix-format features, extracting frequency domain feature components from the physiological signals related to 1-step, 2-step, and 3-step predictions. It captures periodic fluctuation patterns such as circadian rhythms and menstrual cycles hidden in the original time series. Then, it uses filtering methods or embedded methods, or other feature selection methods, to screen the features, selecting the optimal feature subset containing frequency domain periodic components as the second set of selected features. This method compensates for the insufficient expression of periodic information by introducing frequency domain features.

[0091] Step S105b4: The first and second screening features are concatenated using a time series prediction model, and multi-order lagged variables are introduced to incorporate physiological features and subjective perception features within the historical time window into the current moment, thus obtaining an extended feature set.

[0092] This embodiment concatenates the first and second screening features output from steps S105b2 and S105b3 to form a complete feature set containing time-domain and frequency-domain periodic feature components. Based on this, an explicit construction of multi-order lag variables is introduced: for the prediction target at the current moment, the previously screened key features corresponding to several previous time steps are added as new feature columns to the feature set, and physiological and psychological fluctuation information within the historical window is incorporated into the input features at the current moment. After consolidation and the introduction of lag variables, an extended feature set is obtained. This extended feature set is used to subsequently output predicted values ​​of ovarian reserve markers 1, 2, and 3 steps ahead. Here, "step" refers to a preset unit time interval (e.g., one day, one week, or one menstrual cycle), "1 step ahead" means predicting the value at the next time point, "2 steps ahead" means predicting the value at the time point after that, and "3 steps ahead" means predicting the value at the time point after that, corresponding to near-term, mid-term, and slightly longer-term ovarian function estimates, respectively. This extended feature set is used for subsequent prediction steps in a low-redundancy, high-information-density form.

[0093] Step S105b5: Using a time series prediction model, the prediction performance of the extended feature set on multiple prediction tasks is evaluated through statistical tests and multiple cross-validation.

[0094] In this embodiment, the time series prediction model, based on an extended feature set, is evaluated for prediction performance on multiple prediction tasks (e.g., multiple lead-step prediction tasks, including 1-step, 2-step, and 3-step lead-step predictions). The evaluation employs multiple cross-validation (e.g., 10-fold cross-validation, where existing data is divided into several subsets, each subset is used as the validation set, and the remaining subsets are used for model validation; the average of the validation results is taken as the performance metric), and performance metrics (e.g., root mean square error or mean absolute error) under the extended feature set are calculated. Based on this, statistical tests (e.g., paired t-test or Wilcoxon signed-rank test) are performed on the prediction results to verify whether the prediction performance is statistically significant, thus comprehensively evaluating the generalization ability, prediction stability, and result reliability of the extended feature set.

[0095] Step S105b6: Using a time series prediction model, based on an extended feature set, the predicted values ​​of ovarian reserve markers are output through an autoregressive method. The predicted values ​​of ovarian reserve markers are then filtered through multi-criteria decision-making to obtain the ovarian function prediction results.

[0096] In this embodiment, the time series prediction model, based on the extended feature set validated in step S105b5, employs an autoregressive method (e.g., multivariate autoregression) to output predicted values ​​of ovarian reserve markers through linear weighted summation. Building upon this, a multi-criteria decision-making process is introduced, comprehensively considering multiple evaluation criteria such as prediction accuracy, computational cost, and timeliness to filter the output predicted values ​​of ovarian reserve markers and determine the optimal prediction result, which is then output as the ovarian function prediction result. Through this multi-criteria decision-making mechanism, the model can adaptively select the optimal prediction output according to the specific application scenario, balancing prediction accuracy and timeliness.

[0097] This embodiment first converts high-dimensional time-series data into matrix format through feature tiling, eliminating the rigid dependence on the original timestamp order and enabling conventional machine learning algorithms to handle multimodal time-series data, thus reducing modeling complexity. Next, a multi-objective evolutionary feature selection method is used to optimize prediction accuracy, feature quantity, and computational cost, efficiently selecting the optimal subset of time-domain features. Simultaneously, Fourier series transform is used to extract and select frequency-domain periodic features, compensating for the inadequacy of purely time-domain features in representing periodic patterns such as circadian rhythms and menstrual cycles. These two parallel feature selection methods extract key features from both the time and frequency domains, forming a complementary approach. Then, the two types of selected features are combined, and multi-order lag variables are introduced to incorporate physiological and psychological fluctuations within the historical window into the current input, making the extended feature set simultaneously contain short-term dynamics, long-term trends, and periodic patterns. Finally, statistical tests and multiple cross-validation are used to rigorously evaluate the predictive performance of the extended feature set, ensuring the generalization ability, stability, and statistical reliability of the prediction results. Finally, the predicted values ​​of ovarian reserve markers are output through an autoregressive method, and the optimal value is selected by comprehensively considering prediction accuracy, computational cost and timeliness through multi-criteria decision-making. This allows the model to adaptively select the optimal prediction output according to the application scenario, thus balancing prediction accuracy and timeliness.

[0098] In some alternative implementations, the aforementioned multivariate autoregressive model can be replaced with traditional time series statistical models (such as ARMA and ARIMA) or lightweight machine learning regressors (such as XGBoost and LightGBM). The advantages of the ARIMA model are its strong mathematical interpretability, minimal parameter count, and ability to run directly on low-computing-power chips. Tree models such as XGBoost or LightGBM can automatically handle nonlinear relationships and naturally output feature importance during feature selection. Furthermore, the feature selection algorithm can be replaced by principal component analysis for linear dimensionality reduction, which offers advantages such as fast computation speed and fully data-driven operation. Users can choose the appropriate alternative based on their actual deployment scenario (cloud / edge) and accuracy requirements.

[0099] In some optional implementations, the training process for the time series prediction model further includes: Step 1: Obtain the second training dataset, which includes physiological time-series training data of sample users, standard scale rating training data, real-time questionnaire training data, and gold standard label data.

[0100] This embodiment collects a large amount of historical data from users to form the second training dataset. Among these, physiological time-series training data (HRV, body temperature, sleep, etc. continuously collected by wearable devices), standard scale scoring training data (periodic scores such as PSQI and SCL-90), and instantaneous questionnaire training data (instantaneous subjective scores triggered by events) are all high-frequency dynamic input features; the gold standard label data (AMH, AFC, FSH values) are sparse clinically measured indicators. Integrating these data into the training set provides a data foundation for subsequently learning the mapping relationship of ovarian reserve markers from multimodal features.

[0101] Step 2: Tiling the features of the multimodal data in the second training dataset to obtain training features in matrix format.

[0102] This embodiment performs a feature tiling operation on the multimodal temporal feature vector set in the training data, expanding or flattening it along the time axis, and converting it into a two-dimensional matrix format that can be processed by conventional machine learning algorithms. This eliminates the rigid dependence of the input data on the original timestamp order and provides standardized input for subsequent feature selection and model training.

[0103] Step 3: Use a multi-objective evolutionary feature selection method to filter the matrix-format training features to obtain the first candidate feature subset; use Fourier series transform to filter the matrix-format training features to obtain the second candidate feature subset.

[0104] This embodiment employs two parallel methods for feature selection of the trained features in the tiled matrix format. Method one is a multi-objective evolutionary feature selection method: each candidate feature subset is encoded as an individual, and multiple conflicting optimization objectives—maximizing prediction accuracy, minimizing the number of features, and minimizing computational cost—are used. Through iterative evolutionary search (including genetic operations such as selection, crossover, and mutation), a set of Pareto-optimal candidate feature subsets is generated. Method two is a Fourier series transform: the training features are transformed in the frequency domain to extract frequency domain feature components from physiological signals, capturing periodic fluctuation patterns such as circadian rhythms and menstrual cycles. Other feature selection methods, such as filtering or embedded methods, are then used to select the features. The two methods output the first and second candidate feature subsets, respectively.

[0105] Step four: Concatenate the first candidate feature subset with the second candidate feature subset, and introduce multi-order lag variables to construct an expanded training feature set.

[0106] This embodiment concatenates the first and second candidate feature subsets to form a complete feature set containing both time-domain and frequency-domain periodic feature components. Based on this, multi-order lag variables are explicitly constructed: for the prediction target at the current moment, the selected key features corresponding to the previous several time steps are added as new feature columns to the feature set, and physiological and psychological fluctuation information within the historical window is incorporated into the input features at the current moment. Through consolidation and the introduction of lag variables, an information-rich extended training feature set is constructed.

[0107] Step 5: Evaluate the predictive performance of the expanded training feature set using multiple cross-validation and statistical tests.

[0108] This embodiment performs multiple cross-validation on the expanded training feature set (e.g., 10-fold cross-validation, where the training data is divided into several subsets, each subset is used as the validation set, and the remaining subsets are used for training; the average of the validation results is taken as the performance metric). Performance metrics (e.g., root mean square error or mean absolute error) are calculated for the expanded training feature set. Based on this, statistical tests (e.g., paired t-test or Wilcoxon signed-rank test) are performed on the prediction results to verify whether the prediction performance is statistically significant, thereby evaluating the generalization ability and prediction stability of the expanded training feature set.

[0109] Step 6: Based on the cross-validation results, determine the optimal extended feature set configuration through multi-criteria decision-making. Based on the optimal extended feature set configuration, use the regression method to train the model until the model converges, thus obtaining the time series prediction model.

[0110] Based on the cross-validation and statistical test results in step five, this embodiment introduces a multi-criteria decision-making process. It comprehensively considers the performance of the extended training feature set across multiple evaluation criteria, such as prediction accuracy, feature dimensionality, and computational cost, to determine the optimal extended feature set configuration (e.g., determining the optimal combination of lag variable orders or feature pooling weights). Subsequently, an initial time series prediction model is trained based on the optimal configuration. A regression method (e.g., multivariate autoregression) is used to learn the mapping relationship between features and the gold standard label through linear weighted summation. The root mean square error or mean absolute error is used as the loss function, and the model parameters are optimized through backpropagation or least squares until the model converges, resulting in the time series prediction model.

[0111] This embodiment first collects a dataset containing high-frequency multimodal features and sparse gold standard labels, providing a comprehensive data foundation for learning the mapping between physiological and psychological dynamics and ovarian reserve. Secondly, by tiling the training data, the high-dimensional time-series data is converted into a standardized matrix format, eliminating the rigid dependence on the original timestamp order and providing an efficient and unified data input foundation for subsequent feature selection and model training. Next, two parallel feature selection methods—multi-objective evolutionary feature selection and Fourier series transformation—are employed to extract key features from both the time and frequency domains. Time-domain selection achieves optimal balance between prediction accuracy, feature quantity, and computational cost through multi-objective optimization, while frequency-domain selection compensates for the inadequacy of time-domain features in representing circadian rhythms, menstrual cycles, and other periodic patterns through periodic component extraction. These two methods effectively complement each other. Then, by combining multiple feature selection algorithms with cross-validation and multi-criteria decision-making, redundant and noisy features are effectively eliminated, reducing model complexity, improving generalization ability and training efficiency, while preserving feature interpretability. Finally, a multivariate autoregressive output layer is trained based on the optimal feature subset, achieving a linear mapping from selected features to ovarian reserve biomarkers. Finally, the model parameters are optimized through backpropagation of the root mean square error to ensure that the model achieves high-precision and low-error prediction performance on both the training and validation sets. This embodiment lays a solid foundation for lightweight and highly interpretable prediction models, and is particularly suitable for edge deployment and real-time ovarian function prediction scenarios.

[0112] like Figure 2 As shown in the figure, the ovarian function prediction method based on multimodal data provided in this embodiment also includes: Step S106: Use a time series prediction model to perform risk analysis on the ovarian function prediction results to obtain risk classification results and fertility index. The risk classification results indicate the risk level of the user's ovarian function, and the fertility index is used to provide the user's fertility score.

[0113] This embodiment uses the predicted ovarian reserve marker values ​​(such as AMH or AFC values) obtained in step S105 as input. It employs preset risk mapping rules (such as population percentile-based classification or using clustering models to stratify the historical feature space) to convert continuous values ​​into discrete risk classification results, including normal ovarian function, risk of diminished ovarian function, or early warning of premature ovarian insufficiency (POI). Simultaneously, it generates a fertility index score of 0-100 to intuitively quantify the user's current fertility potential. This risk analysis process does not require users to have medical expertise to understand their ovarian status, enabling non-invasive, home-based, and low-cost dynamic monitoring. It can promptly capture subtle changes in ovarian function, providing a crucial time window for egg freezing, conception preparation, or early medical intervention.

[0114] Alternatively, a simple alert indicating an increase or decrease in status could be output. However, this embodiment significantly improves the operability and clinical guidance value of the assessment results by refining the risk level and continuous index score.

[0115] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0116] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and all such modifications and variations fall within the scope defined by the appended claims.

Claims

1. A method for predicting ovarian function based on multimodal data, characterized in that, The method includes: The user's physiological time-series data is collected through wearable devices, and the standard scale score data obtained by the user by filling in a standard scale is acquired. Anomaly detection is performed on the physiological time series data to trigger the collection of real-time questionnaire data, which is used to reflect the instantaneous state of the user when the physiological time series data is abnormal. Feature extraction is performed on the physiological time-series data to obtain physiological features, and feature extraction is performed on the standard scale scoring data and the real-time questionnaire data to obtain subjective perception features; The physiological features and the subjective perception features are fused to obtain a high-dimensional temporal feature vector set; Ovarian function is predicted by using a time series prediction model on the high-dimensional time series feature vector set, and the ovarian function prediction results are obtained.

2. The method according to claim 1, characterized in that, The step of detecting anomalies in the physiological time-series data and triggering the collection of immediate questionnaire data includes: Obtain the user's dynamic personalization baseline; Anomaly detection is performed on the physiological time series data based on the dynamic personalized baseline. The anomaly detection includes a first anomaly detection condition and a second anomaly detection condition. The first anomaly detection condition is to detect whether the proportion of each data in the physiological time series data deviating from the dynamic personalized baseline within a preset time exceeds a preset absolute threshold. The second anomaly detection condition is to detect whether the temporal trend slope of each data in the physiological time series data within a preset time window exceeds the historical normal fluctuation rate. If any data in the physiological time series data satisfies at least one of the first anomaly detection condition and the second anomaly detection condition, then the collection of real-time questionnaire data is triggered.

3. The method according to claim 1, characterized in that, The physiological time-series data includes resting heart rate, heart rate variability data, body temperature data, blood oxygen saturation data, body movement data, and sleep stage data; The step of extracting features from the physiological time-series data to obtain physiological features includes: Feature extraction was performed on the resting heart rate data to obtain the mean resting heart rate at night and the mean resting heart rate during the day; Feature extraction is performed on the heart rate variability data to obtain autonomic nervous system regulation features, which include time-domain features and frequency-domain features. Feature extraction is performed on the body temperature data to obtain circadian rhythm features, which include the diurnal amplitude, diurnal phase, and time of the nighttime minimum of body temperature. Feature extraction was performed on the blood oxygen saturation data to obtain the lowest blood oxygen value at night and the number of blood oxygen drop events; Feature extraction is performed on the physical activity data to obtain the daytime activity level and activity intensity; Feature extraction was performed on the sleep stage data to obtain the duration of deep sleep, the duration of rapid eye movement sleep, the duration of light sleep, and the number of nighttime awakenings.

4. The method according to claim 1, characterized in that, The process of extracting features from the standard scale scoring data and the immediate questionnaire data to obtain subjective perception features includes: Feature extraction is performed on the standard scale scoring data to obtain long-term statistical features, which include the average value and volatility of the standard scale scoring data over a preset long period. Feature extraction is performed on the real-time questionnaire data to obtain short-term dynamic features, which include the average value and volatility of the real-time questionnaire scores within a preset short period, as well as the cross-correlation coefficient between the real-time questionnaire scores and the intensity of physiological events. The subjective perception features are obtained by combining the long-term statistical features and the short-term dynamic features.

5. The method according to claim 1, characterized in that, The step of using a time series prediction model to predict ovarian function from the high-dimensional time-series feature vector set, and obtaining the ovarian function prediction results, includes: The high-dimensional time-series feature vector set is augmented with positional encoding using a time-series prediction model to obtain a time-series feature sequence with positional encoding. The time series feature sequence with position encoding is processed in parallel using the self-attention mechanism of the time series prediction model to capture the long-term dependencies between different time steps and the cross-modal associations between the physiological features and the subjective perception features, thereby obtaining a deep time series feature representation. A time series prediction model is used to linearly map the deep time series feature representation to obtain the predicted values ​​of ovarian reserve markers, which are used as the ovarian function prediction results.

6. The method according to claim 5, characterized in that, The training process of the time series prediction model includes: Obtain a first training dataset, which includes physiological time-series training data of sample users, standard scale rating training data, real-time questionnaire training data, and gold standard label data. Align the multimodal data in the first training dataset according to the time window to obtain the aligned multimodal temporal feature vector set; Positional encoding is added to the multimodal time series feature vector set, and the initial time series prediction model is trained by parallel processing and linear mapping through self-attention mechanism to obtain the first predicted value; The root mean square error is calculated on the first predicted value and the gold standard label data to obtain the first root mean square error result. Based on the first root mean square error result, the model parameters are optimized through backpropagation until the model converges, thus obtaining the time series prediction model.

7. The method according to claim 1, characterized in that, The step of using a time series prediction model to predict ovarian function from the high-dimensional time series feature vector set to obtain ovarian function prediction results also includes: The high-dimensional time-series feature vector set is tiled using a time series prediction model to obtain matrix-format features. The matrix-formatted features are filtered using a multi-objective evolutionary feature selection method based on a time series prediction model to obtain the first filtered features. The matrix-formatted features are filtered using a time series prediction model and Fourier series transform to obtain the second filtered features. The first and second selected features are concatenated using a time series prediction model, and multi-order lagged variables are introduced to incorporate physiological features and subjective perception features within the historical time window into the current moment, thus obtaining an extended feature set. The predictive performance of the extended feature set on multiple prediction tasks was evaluated using a time series prediction model through statistical tests and multiple cross-validation. Using a time series prediction model, the predicted values ​​of ovarian reserve markers are output through an autoregressive method based on the extended feature set. The predicted values ​​of ovarian reserve markers are then filtered through multi-criteria decision-making to obtain the predicted ovarian function results.

8. The method according to claim 7, characterized in that, The training process of the time series prediction model also includes: Obtain a second training dataset, which includes physiological time-series training data of sample users, standard scale rating training data, real-time questionnaire training data, and gold standard label data. The multimodal data in the second training dataset are tiled to obtain training features in matrix format; A multi-objective evolutionary feature selection method is used to filter the matrix-format training features to obtain a first candidate feature subset; a Fourier series transform is used to filter the matrix-format training features to obtain a second candidate feature subset. The first candidate feature subset and the second candidate feature subset are concatenated, and multi-order lag variables are introduced to construct an expanded training feature set; The predictive performance of the expanded training feature set was evaluated using multiple cross-validation and statistical tests. Based on the cross-validation results, the optimal extended feature set configuration is determined through multi-criteria decision-making. The model is then trained using a regression method based on the optimal extended feature set configuration until the model converges, thus obtaining the time series prediction model.

9. The method according to claim 1, characterized in that, It also includes: using the time series prediction model to perform risk analysis on the ovarian function prediction results to obtain risk classification results and fertility index, wherein the risk classification results indicate the risk level of the user's ovarian function, and the fertility index is used to provide the user's fertility score.

10. An ovarian function prediction device based on multimodal data, characterized in that, include: A processor and a memory connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to cause the processor to perform the ovarian function prediction method based on multimodal data as described in any one of claims 1-9.

Images