A method for predicting the risk of severe illness based on the PINN-Cox joint model and related equipment.
By combining the PINN-Cox joint model with deep learning and disease evolution mechanisms, the problems of traditional models' difficulty in fitting complex severe cases and overfitting of deep learning models are solved, achieving high accuracy and interpretability in survival prediction for critically ill patients.
Patent Information
- Application Number
- CN202610905829.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-23
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2046-06-23
AI Technical Summary
In the existing fields of intensive care and cancer treatment, traditional Cox proportional hazards models are difficult to fit the complex nonlinear interactions of critical illnesses. Deep learning models overfit under small sample and noisy conditions and lack interpretability, resulting in inaccurate and unstable survival predictions for critically ill patients.
The PINN-Cox joint model is adopted. By constructing a deep neural network and introducing a non-proportional Cox model and a physical residual loss term from the differential equation of disease evolution mechanism, a mechanism-data joint loss function is defined, joint constraint training is performed, and a time-dependent relative risk function is output.
It achieves high-precision, robust, and medically interpretable dynamic prediction of survival prognosis for critically ill patients under conditions of small sample size and incomplete clinical data, and provides accurate predictions of individualized survival curves and expected remaining lifespan.
Smart Images

Figure CN122658670B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent medical information technology, and in particular to a method for predicting the risk of severe illness based on the PINN-Cox joint model and related equipment. Background Technology
[0002] In the fields of intensive care and oncology treatment, accurately predicting patient disease progression and survival prognosis is crucial for clinical decision-making. Current clinical prediction mainly relies on two types of methods: one is statistical methods, represented by the traditional Cox proportional hazards model. Although this method has statistical interpretability, it relies on the proportional hazards assumption and the linear assumption of covariates. When facing real medical scenarios with highly time-varying characteristics and complex nonlinear interactions, such as tumor growth or severe viral infection, it often fails to accurately fit dynamic risks. The other type is pure data-driven survival prediction models based on deep learning. Although these models overcome linear constraints through complex network architectures and can handle high-dimensional medical data, they are essentially black-box models, completely dependent on a large number of high-quality samples. Under the conditions of small samples, strong noise, and irregular follow-up common in the clinical world, they are prone to overfitting and unstable extrapolation predictions. Furthermore, they lack constraints on the intrinsic pathological evolution mechanisms of diseases, making it difficult for the interpretability and robustness of the models to meet the stringent requirements of clinical medicine. Summary of the Invention
[0003] The purpose of this invention is to provide a critical illness risk prediction method and related equipment based on the PINN-Cox joint model, in order to solve one or more technical problems existing in the prior art, and at least provide a beneficial option or create conditions that enable high-precision, high-robustness and medically interpretable dynamic prediction of the survival prognosis of critically ill patients under conditions of small sample size and incomplete clinical data, by integrating the dual constraints of disease evolution mechanism and data-driven approach.
[0004] On the one hand, this application provides a method for predicting the risk of severe illness based on the PINN-Cox joint model, including the following steps: Acquire multi-source time-series clinical data of patients with target severe diseases, perform structured processing and time alignment, and construct an individual survival analysis dataset; A PINN-Cox joint prediction model is constructed, which inputs time variables and dynamic feature vectors into a deep neural network and outputs a time-dependent relative risk function through nonlinear mapping to characterize the instantaneous hazard rate under the non-proportional Cox model. Define a mechanism-data joint loss function, and use the multi-source time-series clinical data to jointly constrain the training of the PINN-Cox joint prediction model; the mechanism-data joint loss function includes a negative log-partial likelihood loss term based on the non-proportional Cox model, and a physical residual loss term based on the differential equation of the evolution mechanism of severe diseases; Input the time series data of the patient to be predicted into the trained model, and output the patient's survival function, cumulative risk function and median expected life expectancy within a future preset time window.
[0005] Furthermore, the acquisition of multi-source time-series clinical data from patients with the target severe illness, and the subsequent structured processing and time alignment, specifically includes: Longitudinal clinical data of patients are obtained from the hospital's electronic medical record system or intensive care database. The longitudinal clinical data includes etiological and tumor marker indicators, vital signs and physiological and biochemical indicators, imaging follow-up characteristics, and clinical treatment events. The acquired longitudinal clinical data were de-identified, outlier cleaned, unit standardized, and missing value processed. A patient-level timeline is constructed by setting a predetermined time granularity, and observation records of different frequencies are mapped to corresponding time points to achieve time alignment.
[0006] Furthermore, the missing value processing specifically involves: using a combination of nearest-neighbor observation imputation and interpolation estimation based on disease dynamics to complete the missing values and generate missing indicator variables; the method also includes extracting key features characterizing disease evolution and mortality risk, the key features including baseline static features and dynamic evolution features, wherein the dynamic evolution features include indicator levels, rate of change, fluctuation amplitude, and stage mutation information.
[0007] Furthermore, the construction of the PINN-Cox joint prediction model specifically includes: Construct an input layer to receive the time-series follow-up data after feature engineering and the corresponding timestamps; A deep feature extraction network is constructed using the Transformer architecture to learn complex nonlinear representations related to prognosis and survival from high-dimensional features. A time-based modeling and dynamic state layer is constructed, and the derivative of the neural network output with respect to time is calculated using automatic differentiation technology to deduce physically meaningful latent state variables, including inflammation intensity or tumor burden. Construct an output layer to output the time-varying risk function and the instantaneous hazard rate of the non-proportional Cox model.
[0008] Furthermore, the mechanism-data joint loss function is specifically defined as follows: The total loss is equal to the weighted sum of the data fitting term loss, the physical mechanism term loss, and the regularization term loss; Wherein, the data fitting term loss is the Cox average negative log partial likelihood loss based on the sampling risk set approximation; The physical mechanism term loss is the residual of the disease progression differential equation constructed using automatic differentiation, specifically the residual of the dynamic model characterizing the pathological evolution of the target organ, including the residual of the tumor Gompertz growth model or the dynamic residual of the influenza virus. The regularization term loss is an L1 or L2 norm penalty term.
[0009] Furthermore, the joint constraint training of the PINN-Cox joint prediction model using the multi-source time-series clinical data specifically includes: The gradient of the physical residual loss term is calculated using automatic differentiation techniques, and the network parameters are updated synchronously during backpropagation. The evolutionary mechanism of severe diseases is embedded as a physical constraint in the network training process. By dynamically adjusting the weight coefficients to balance the data fitting term and the physical mechanism term, the network is forced to satisfy the consistency of clinical observation data and the consistency of disease evolution mechanism during the optimization process. The physical constraint term is implemented by imposing penalties on solutions that deviate from the evolution law of the differential equation.
[0010] Furthermore, the output of the patient's survival function within a future preset time window specifically includes: Based on the trained model, individualized survival rate-time curves and expected remaining survival are derived. The changes in key potential pathological states constrained by mechanistic equations are simultaneously inverted and output. These key potential pathological states include the in vivo viral load change trajectory or the dynamic evolution trend of tumor burden indicated by mechanistic dynamic residuals. Based on the survival function and the key potential pathological states, clinical risk warning information is generated.
[0011] On the other hand, this application provides a critical illness risk prediction system based on the PINN-Cox joint model, characterized in that it includes a server for data processing and model inference; The server includes a processor, a memory, and a communication interface. The memory stores a computer program, and the communication interface is used to connect to the hospital information system to obtain multi-source time-series data of patients. The processor is used to execute the computer program to implement the critical illness risk prediction method based on the PINN-Cox joint model as described above. The server is also configured with: The mechanism-data dual-driven inference engine is used to load the trained PINN-Cox joint prediction model, perform forward inference on the input patient time-series features, and output individualized survival functions and potential pathological state trajectories. The risk warning and visualization module is used to dynamically generate multi-level risk warning signals based on the declining slope of the survival function and the degree of deterioration of potential pathological conditions, and to render individual survival curves and risk trend graphs on the connected display terminal.
[0012] On the other hand, this application provides an electronic device, including a processor and a memory; The memory is used to store computer programs; The processor is used to execute the computer program stored in the memory to implement the critical illness risk prediction method based on the PINN-Cox joint model as described above.
[0013] On the other hand, this application provides a computer-readable storage medium storing computer program instructions thereon, which, when executed by a processor, implement the aforementioned method for predicting the risk of severe illness based on the PINN-Cox joint model.
[0014] The beneficial effects of this invention are as follows: This application provides a method for predicting the risk of severe illness based on the PINN-Cox joint model. This technology acquires and structures multi-source time-series clinical data to construct an individual survival analysis dataset, providing a high-quality input foundation for the model. On this basis, a PINN-Cox joint prediction model is constructed, which utilizes the nonlinear mapping capability of deep neural networks to output a time-dependent relative risk function, effectively breaking through the proportional risk assumption limitation of the traditional Cox model and accurately representing the instantaneous hazard rate in complex severe illness scenarios. At the same time, this method innovatively defines a mechanism-data joint loss function, introducing a negative log-partial likelihood loss term based on the non-proportional Cox model and a physical residual loss term based on the differential equation of disease evolution mechanism during training. This integrates medical pathology laws as physical constraints into a pure data-driven black box model, significantly improving the robustness and anti-overfitting ability of the model under incomplete clinical data conditions such as small samples and irregular follow-ups. Ultimately, it achieves high-precision and medically interpretable dynamic prediction of the survival function, cumulative risk function, and median expected remaining life of the patients to be predicted within a future preset time window. This application also provides related equipment for the above method. The beneficial effects of the related equipment are similar to those of the above method, and will not be described in detail here.
[0015] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the description, claims and drawings. Attached Figure Description
[0016] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.
[0017] Figure 1 This is a flowchart of the critical illness risk prediction method based on the PINN-Cox joint model provided in this application; Figure 2 This is a schematic diagram of the critical illness risk prediction framework based on the PINN-Cox joint model provided in this application. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0019] The present application will be further described below with reference to the accompanying drawings and specific embodiments. The described embodiments should not be considered as limitations on the present application, and all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of the present application.
[0020] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.
[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.
[0022] In clinical medicine fields such as intensive care and cancer treatment, accurately predicting a patient's disease progression and survival prognosis is crucial for developing personalized treatment plans. With the development of medical informatization, electronic medical records and monitoring devices have accumulated massive amounts of clinical data. How to utilize this data to uncover disease progression patterns has become a research hotspot in biomedical informatics. Especially for critical illnesses with highly nonlinear evolutionary characteristics, such as severe influenza and malignant tumors, clinicians urgently need an intelligent auxiliary tool capable of processing multi-source heterogeneous data and possessing good interpretability to address the problems of high subjectivity and low accuracy in traditional manual assessments.
[0023] Current clinical prognostic prediction technologies can be broadly categorized into two types. The first type is survival analysis methods based on traditional statistics, with the Cox proportional hazards model being the most representative. This method assesses the impact of different factors on patient survival time by establishing a regression relationship between covariates and a hazard function. The second type comprises pure data-driven prediction models based on deep learning, which have emerged in recent years. These methods typically utilize recurrent neural networks, long short-term memory networks, or Transformer architectures to directly learn features from time-series data such as electronic medical records. They attempt to fit the nonlinear relationships in the data through complex network structures, thereby predicting the patient's probability of survival or remaining lifespan.
[0024] However, the aforementioned existing technologies all have significant shortcomings and deficiencies in practical applications. Traditional statistical methods, such as the Cox proportional hazards model, rely on the proportional hazards assumption and the linear covariate assumption, assuming that the hazard ratio remains constant over time and that the variables have a linear additive relationship. However, in real-world medical scenarios such as tumor growth or severe viral infections, disease evolution exhibits highly time-varying characteristics and complex nonlinear interactions. Traditional models struggle to accurately fit this dynamic risk, resulting in limited predictive accuracy.
[0025] While deep learning-based, purely data-driven models overcome linear constraints with their powerful fitting capabilities, they remain essentially black-box models. These models rely entirely on training with a large number of high-quality labeled samples. However, in the real world of clinical practice, critical care data is often characterized by small sample sizes, high noise levels, and irregular follow-up records. When data is scarce or of low quality, purely data-driven models are highly susceptible to overfitting, leading to highly unstable extrapolation predictions. Furthermore, lacking constraints on the intrinsic pathological evolution mechanisms of diseases, these models cannot provide explanations consistent with medical logic, and their robustness and interpretability fail to meet stringent clinical requirements.
[0026] To address the aforementioned issues, this application proposes a method and related equipment for predicting the risk of severe illness based on a PINN-Cox joint model. This application innovatively embeds the differential equations of the medical mechanisms underlying the evolution of severe diseases as physical constraints into a deep learning network by constructing a PINN-Cox joint prediction model. It utilizes a mechanism-data joint loss function, including negative log-partial likelihood loss and physical residual loss, for joint constraint training. This method not only overcomes the proportional hazards assumption limitations of traditional Cox models by leveraging the nonlinear mapping capabilities of neural networks, but also solves the problems of overfitting and lack of interpretability in purely data-driven models under conditions of small samples and incomplete clinical data by introducing a physical residual loss term. Thus, it achieves high-precision, robust, and medically interpretable dynamic prediction of the survival prognosis of critically ill patients without requiring a large amount of high-quality labeled data.
[0027] Furthermore, in this application, PINN (Physical Information Neural Network) breaks through the black-box limitations of traditional deep learning in the medical field by embedding the objective medical laws governing the evolution of severe diseases as a strongly constrained model. It calculates the residuals of differential equations using automatic differentiation techniques, forcing the network to follow pathophysiological laws (such as tumor growth or viral dynamics) while learning from massive amounts of clinical data. This addresses the problems of sparsity and noise interference in clinical data, while endowing the model with the ability to invert key potential pathological states, greatly improving the scientific rigor and interpretability of predictions.
[0028] The Cox proportional hazards model in this application provides a rigorous statistical framework for survival analysis, enabling precise quantification of the dynamic relationship between multi-source time-series features and patient prognostic outcomes. It transforms the high-dimensional, complex features extracted from deep networks into time-varying risk functions and instantaneous hazard rates with clear clinical guidance value, overcoming the bottleneck of traditional models' inability to handle non-equidistant sampling data. This allows the system to output individualized survival curves and expected remaining survival in real time, providing a reliable quantitative basis for tiered diagnosis and timely intervention in critically ill patients.
[0029] First, the critical illness risk prediction method based on the PINN-Cox joint model provided in this application will be described in detail below with reference to the accompanying drawings.
[0030] Reference Figure 1 The implementation of the critical illness risk prediction method based on the PINN-Cox joint model provided in this application includes, but is not limited to, the following steps.
[0031] Step S100: Obtain multi-source time-series clinical data of the target critically ill patients, perform structured processing and time alignment, and construct an individual survival analysis dataset.
[0032] Step S100 lays a solid and standardized data foundation for subsequent complex model calculations. In real clinical scenarios, patient medical data often originates from multiple heterogeneous systems such as electronic medical records, monitoring equipment, and laboratory reports, exhibiting fragmented characteristics of multiple sources, high frequency, and varying sampling intervals. This step, through rigorous structuring and time-alignment of these raw, multi-source time-series clinical data, effectively eliminates data misalignment and noise interference in the time dimension, transforming messy clinical records into a standardized individual survival analysis dataset readable by computer models. This ensures that the data input to the model retains the dynamic evolution details of the patient's condition while meeting the stringent requirements of deep learning algorithms for tensor dimensions.
[0033] Step S200: Construct the PINN-Cox joint prediction model, input the time variable and dynamic feature vector into the deep neural network, and output the time-dependent relative risk function through nonlinear mapping to characterize the instantaneous hazard rate under the non-proportional Cox model.
[0034] Step S200 breaks through the rigid assumptions of traditional statistical models regarding data distribution. By constructing a PINN-Cox joint prediction model and inputting time variables and dynamic feature vectors into a deep neural network, this step utilizes the powerful nonlinear mapping capability of neural networks. Instead of forcibly assuming a constant risk ratio, it directly fits a relative risk function that dynamically changes over time.
[0035] This design enables the model to keenly capture the complex nonlinear interactions and time-varying characteristics in the development of severe diseases, accurately characterize the instantaneous hazard rate under the assumption of a non-proportional Cox model, thereby significantly improving the model's fitting accuracy and expressive flexibility when facing complex pathological evolution.
[0036] Step S300: Define the mechanism-data joint loss function and perform joint constraint training on the PINN-Cox joint prediction model using multi-source time-series clinical data. The mechanism-data joint loss function includes a negative log-partial likelihood loss term based on the non-proportional Cox model and a physical residual loss term based on the differential equation of the evolution mechanism of severe diseases.
[0037] In step S300, a deep fusion of medical prior knowledge and clinical observation data is achieved by defining a joint mechanism-data loss function, which is key to solving the problems of small sample size and high noise. This step introduces dual constraints during training: on the one hand, it uses a negative log-partial likelihood loss term based on a non-proportional Cox model to mine statistical regularities in the data; on the other hand, it creatively introduces a physical residual loss term based on the differential equation of the evolution mechanism of severe diseases. This physical constraint term forces the output of the neural network to conform to the objective biological laws of disease development, which is equivalent to equipping the black box model with medical logic navigation, effectively preventing overfitting when the data is scarce or of poor quality, and greatly enhancing the robustness and generalization ability of the model in real clinical environments.
[0038] Step S400: Input the time series data of the patient to be predicted into the trained model, and output the patient's survival function, cumulative risk function and median expected remaining life within a future preset time window.
[0039] In step S400, clinicians are provided with intuitive, quantitative decision support with a time dimension. By inputting real-time time-series data of the patient to be predicted into a model trained with mechanistic constraints, the system can output the patient's survival function, cumulative risk function, and median expected life expectancy within a preset future time window.
[0040] This process not only transforms complex model calculations into clinically understandable prognostic indicators, but also helps doctors dynamically assess patients' future risk trends, thereby providing scientific and precise quantitative basis for developing personalized critical care treatment plans, adjusting nursing strategies, and allocating medical resources.
[0041] In some embodiments of this application, step S100 involves acquiring multi-source time-series clinical data of the target critically ill patient and performing structured processing and time alignment, specifically including: Step S110: Obtain the patient's longitudinal clinical data from the hospital's electronic medical record system or intensive care database. The longitudinal clinical data includes etiological and tumor marker indicators, vital signs and physiological and biochemical indicators, imaging follow-up characteristics, and clinical treatment events.
[0042] In step S110, the complex pathophysiological state of critically ill patients is captured comprehensively and from multiple dimensions, providing rich information input for the model. Because the evolution of critical illnesses such as malignant tumors or severe influenza involves the interaction of multiple systems throughout the body, a single data source cannot fully depict the entire picture of the disease.
[0043] This step extracts longitudinal clinical data from hospital electronic medical record systems or intensive care unit databases, encompassing etiological and tumor markers reflecting disease-specific progression, vital signs and physiological and biochemical indicators characterizing the body's immediate physiological state, imaging follow-up features reflecting macroscopic lesion changes, and clinical treatment events that interfere with disease progression. This integration of multi-source data ensures that the model can simultaneously detect fluctuations in patients' microscopic biochemical indicators and changes in macroscopic clinical symptoms, laying a solid data foundation for subsequent accurate predictions of survival risk.
[0044] Step S120 involves de-identifying, cleaning outliers, standardizing units, and handling missing values in the acquired longitudinal clinical data.
[0045] In step S120, the quality and security of the input model data are ensured. Raw clinical data often contains patient privacy information, abnormal noise caused by equipment malfunctions, and inconsistent units due to different human recording habits. By performing de-identification, this step strictly guarantees patient privacy compliance; outlier cleaning and unit standardization eliminate data bias caused by measurement errors or inconsistent standards; and missing value handling, addressing common clinical missed detections, effectively fills information gaps.
[0046] This series of preprocessing measures transforms non-standardized raw records into high-quality tensors that computers can understand, preventing low-quality data from misleading or interfering with the training of deep learning models.
[0047] Step S130: Set a predetermined time granularity to construct a patient-level time axis, and map observation records of different frequencies to corresponding time points to achieve time alignment.
[0048] Step S130 addresses the challenge of inconsistent sampling frequencies across multiple medical data sources over time, a crucial step in achieving temporal feature fusion. In clinical practice, vital signs may be recorded every minute, while biochemical indicators may be recorded daily, and imaging examinations may only be performed every few weeks. This asynchronous sampling frequency leads to data misalignment along the timeline.
[0049] This step constructs a unified patient-level timeline by setting a predetermined time granularity, and maps observation records of different frequencies to corresponding time points through mathematical methods such as interpolation or forward padding, achieving strict alignment of multimodal data in the time dimension. This enables the model to simultaneously read all physiological, biochemical, and imaging features of the patient at each unified time step, thereby accurately capturing the dynamic coupling relationship between different indicators over time.
[0050] In some embodiments of this application, in step S120, the missing value processing specifically involves: completing the missing value by combining neighbor observation imputation with interpolation estimation based on disease dynamics and trends, and generating a missing indicator variable.
[0051] In step S120, missing value handling is used to overcome the shortcomings of traditional simple imputation methods that ignore the temporal correlation and pathophysiological laws of medical data, thereby maximizing information utilization while preserving the true distribution characteristics of the data. In critical care settings, missing data is often not random, and the changes in different indicators follow specific biological trajectories. By combining nearest-neighbor imputation with interpolation estimation based on disease dynamics, this method can not only maintain the short-term continuity of data by utilizing temporally adjacent effective observations, but also infer the numerical trend over long intervals based on the intrinsic dynamic evolution of specific critical diseases (such as tumor growth or viral load changes).
[0052] This mechanism-based interpolation avoids the abrupt changes that linear interpolation might introduce, which violate medical common sense, making the completed data more consistent with the actual progression of the patient's condition. Simultaneously, generating missing indicator variables and inputting them as additional features into the model allows the model to clearly distinguish between observed and estimated data, effectively differentiating data credibility and preventing noise introduced by imputation from misleading model training. This significantly improves the model's robustness and predictive accuracy when faced with incomplete clinical data.
[0053] In some embodiments of this application, the above method further includes extracting key features characterizing disease evolution and mortality risk. The key features include baseline static features and dynamic evolution features, wherein the dynamic evolution features include indicator levels, rates of change, fluctuation amplitudes, and stage mutation information.
[0054] Specifically, a high-dimensional feature space is constructed that can comprehensively map the complex evolutionary patterns of severe diseases. By deeply mining the clinical semantics behind the data, the model's ability to discriminate mortality risk is significantly improved. This step is no longer limited to numerical records at a single point in time, but subdivides the feature dimensions into baseline static features and dynamic evolutionary features. The baseline static features lock in relatively fixed background information such as the patient's age, gender, and underlying medical history, establishing an initial benchmark for individual risk. The dynamic evolutionary features are further refined into indicator levels, rates of change, fluctuation amplitudes, and information on stage-specific mutations.
[0055] The indicator levels reflect the patient's current physiological state and disease severity, the rate of change captures the instantaneous speed of disease deterioration or improvement, the fluctuation amplitude reflects the stability and compensatory capacity of the body's physiological functions, and the stage mutation information accurately identifies key clinical events such as explosive tumor growth or severe disease outcomes. This refined feature engineering enables the model not only to perceive the patient's current state but also to understand the trend and potential risks of disease development, thus providing highly discriminative information support for subsequent high-precision survival prediction.
[0056] In some embodiments of this application, step S200, constructing the PINN-Cox joint prediction model, specifically includes: Step S210: Construct an input layer to receive time-series follow-up data processed by feature engineering and the corresponding timestamps.
[0057] In step S210, a standardized connection is established between physical world clinical records and digital spatial computing models. The input layer constructed in this step is specifically designed to receive time-series follow-up data that has undergone complex feature engineering processing, along with its strictly corresponding timestamps. This is not merely a simple data transfer, but rather to ensure that the model can accurately perceive the precise location of each clinical observation in the patient's life course.
[0058] By receiving feature vectors containing the time dimension, the input layer provides the basic coordinates for the subsequent network to understand the dynamic evolution of the disease. This ensures that the model can maintain the continuity of temporal logic when processing critical clinical data with non-equal interval sampling, laying a solid physical foundation for deep networks to capture temporal dependencies.
[0059] Step S220: Construct a deep feature extraction network using the Transformer architecture to learn complex nonlinear representations related to prognosis and survival from high-dimensional features.
[0060] In step S220, a deep feature extraction network is constructed using the advanced Transformer architecture, overcoming the limitations of traditional recurrent neural networks in processing long-sequence medical data and achieving accurate capture of the deep semantics of high-dimensional clinical features. The prognosis of critically ill patients often depends on the complex nonlinear interactions between indicators of multiple organ systems, rather than the linear superposition of a single indicator.
[0061] This step, through the self-attention mechanism unique to the Transformer architecture, can process high-dimensional input features in parallel, automatically weighting and focusing on key indicators that have a decisive impact on survival prognosis while ignoring irrelevant noise. It learns complex nonlinear representations related to prognosis and survival from high-dimensional features, effectively uncovering pathological correlation patterns hidden behind massive amounts of clinical data, and providing highly discriminative high-order abstract features for subsequent risk calculations.
[0062] Step S230: Construct a time modeling and dynamic state layer, use automatic differentiation technology to calculate the derivative of the neural network output with respect to time, and deduce the potential state variables with physical meaning, including inflammation intensity or tumor burden.
[0063] In step S230, the output of the black-box neural network is transformed into a pathophysiological state with a clear medical interpretation, achieving a deep integration of data-driven and mechanistic models. This step constructs a time modeling and dynamic state layer, and uses automatic differentiation technology to accurately calculate the derivative of the neural network output with respect to time. This is equivalent to teaching the model calculus, enabling it to infer the underlying dynamic mechanism based on the changing trends of the data.
[0064] Through this process, the model can extrapolate physically significant latent state variables, such as inflammation intensity or tumor burden. While these variables are difficult to monitor continuously in real-time during routine clinical examinations, they are the core drivers determining the life and death of critically ill patients. By making these latent variables explicit through mathematical deduction, the model is no longer merely fitting data, but simulating the actual evolution of disease within the patient's body, greatly enhancing the model's interpretability and scientific rigor.
[0065] Step S240: Construct the output layer to output the time-varying risk function and the instantaneous hazard rate of the non-proportional Cox model.
[0066] In step S240, the abstract features extracted by the deep network and the physical states deduced by the dynamics layer are transformed into quantitative risk indicators that clinicians can directly use. This step constructs an output layer specifically for outputting the time-varying risk function and the instantaneous hazard rate of the non-proportional Cox model.
[0067] This design breaks free from the traditional Cox model's assumption of a constant risk ratio, allowing the model to adjust its mortality risk predictions in real time based on the patient's dynamic characteristics and potential pathological states over time. The output instantaneous hazard ratio accurately characterizes the severity of a patient's condition at a specific point in time, capturing transient fluctuations in disease progression or remission. This not only provides a foundation for calculating the survival function but also allows physicians to visualize the trajectory of risk over time, thus providing a dynamic basis for developing phased precision treatment strategies.
[0068] In some embodiments of this application, in step S300, the mechanism-data joint loss function is specifically defined as: the total loss is equal to the weighted sum of the data fitting term loss, the physical mechanism term loss, and the regularization term loss.
[0069] Among them, the data fitting term loss ensures that the model can accurately capture and fit the distribution characteristics of multi-source time-series clinical data in the real world; the physical mechanism term loss embeds the evolution law of critical illness as a strong constraint in the form of differential equation residuals, forcing the network output to conform to the objective biological development logic, thereby providing reliable prior guidance when the data is sparse or noisy; and the regularization term loss is used to penalize overly complex network parameters, effectively preventing the model from overfitting to the training set.
[0070] This weighted fusion mechanism not only achieves deep complementarity between medical mechanisms and massive observational data, but also ensures that the model has strong robustness and interpretability while pursuing high accuracy, enabling it to stably output predictive results that conform to medical common sense in harsh clinical environments.
[0071] In some embodiments of this application, the data fitting term loss is the Cox mean negative log-partial likelihood loss based on the sampling risk set approximation, which is used to solve the computational bottleneck faced by traditional survival analysis models when processing large-scale critical clinical data, while ensuring that the model can accurately capture the statistical regularities in the data.
[0072] Specifically, in high-frequency sampling scenarios such as intensive care, the number of patients is enormous and the follow-up period is intensive. If the standard Cox model is used to calculate the partial likelihood, a risk set needs to be constructed in the entire population, which will result in extremely high computational complexity. By introducing a sampling risk set approximation technique, this step can significantly reduce the amount of computation while maintaining the unbiasedness of statistical inference, making it possible to train the model on massive amounts of data.
[0073] Meanwhile, as a negative log-partial likelihood loss, it drives the model to learn to distinguish the risk differences among different patients at the same time, that is, to maximize the likelihood probability of the observed data, ensuring that the model's prediction results can best fit the actual observed survival outcomes statistically, and is the data foundation for the model to achieve high-precision prediction.
[0074] In some embodiments of this application, the physical mechanism term loss is the residual of the disease progression differential equation constructed using automatic differentiation, specifically the residual of the dynamic model characterizing the pathological evolution of the target organ, including the residual of the tumor Gompertz growth model or the residual of the influenza virus dynamics.
[0075] Specifically, prior knowledge from the medical field is embedded as a strong constraint into the deep learning model, ensuring that the model's predictions not only statistically fit the data but also conform to objective laws in a biological sense. For the complex pathological evolution of severe diseases, this step no longer relies on blind fitting by black-box neural networks but directly introduces dynamic models characterizing the pathological evolution of target organs, such as the residuals of the Gompertz growth model describing the S-shaped growth of tumor volume over time, or the dynamic residuals of influenza virus characterizing the interaction between viral load and immune cells.
[0076] By calculating the deviation of the network output from these differential equations (i.e., residuals) using automatic differentiation techniques and backpropagating this residual as part of the loss function, the model is forced to learn feature trajectories that must adhere to these classical medical physics laws. This mechanism effectively prevents the model from making predictions that violate medical common sense under conditions of sparse data or noise interference (such as tumors growing indefinitely or viruses disappearing in violation of dynamic laws), greatly improving the model's interpretability, robustness, and generalization ability in small sample situations.
[0077] In some embodiments of this application, the regularization loss is an L1 or L2 norm penalty term, which effectively prevents the PINN-Cox joint prediction model from overfitting when faced with high-dimensional, sparse and noisy critical clinical data.
[0078] Specifically, during the training of deep learning networks, models often tend to force the memorization of random noise or specific details in the training data by increasing the weight parameters, which leads to a decline in performance on unknown test sets. By introducing L1 or L2 norms as penalty terms, this step imposes constraints on excessively large weight parameters in the loss function, forcing the network to seek a smoother and simpler solution space during the optimization process.
[0079] The L1 norm tends to produce sparse weights, which helps in feature selection and eliminates redundant clinical indicators. The L2 norm, on the other hand, limits the size of the weights, preventing the model from becoming overly sensitive to a single feature and enhancing model stability. This mechanism ensures that the model not only focuses on improving prediction accuracy but also emphasizes the universality and robustness of the learned pathological features, thus maintaining reliable predictive performance even when facing new patient populations.
[0080] In some embodiments of this application, step S300, which involves jointly training the PINN-Cox joint prediction model using multi-source time-series clinical data, specifically includes: Step S310: Calculate the gradient of the physical residual loss term using automatic differentiation and update the network parameters synchronously during backpropagation.
[0081] In step S310, an efficient mathematical optimization path is constructed to solve the complex differential equation solving and parameter updating challenges in the PINN-Cox model. During the training process of the PINN-Cox model, the physical residual loss term involves calculating the higher-order derivatives of the network output with respect to time or space. Traditional numerical differentiation methods often suffer from high computational cost and limited accuracy. This step utilizes automatic differentiation techniques to accurately and efficiently calculate the gradient of the physical residual loss term relative to the network weights, without relying on approximation errors.
[0082] By synchronously updating network parameters during backpropagation, this step ensures that the model can be precisely corrected along the negative direction of the gradient in each iteration. This minimizes not only the difference between the predicted results and the true labels, but also the deviation between the predicted trajectory and physical laws. This mechanism guarantees that physical constraints can directly drive the evolution of the network, just like data labels, enabling the model to quickly converge in the parameter space to the optimal solution region that conforms to both statistical and physical laws.
[0083] Step S320: The evolutionary mechanism of severe diseases is embedded as a physical constraint in the network training process. By dynamically adjusting the weight coefficients to balance the data fitting term and the physical mechanism term, the network is forced to simultaneously satisfy the consistency of clinical observation data and the consistency of disease evolution mechanism during the optimization process.
[0084] Step S320 addresses the contradiction between the poor interpretability of purely data-driven models in medical scenarios and the weak adaptability of purely mechanistic models. This step transforms the evolutionary mechanisms of severe diseases, such as tumor growth dynamics or viral replication patterns, into specific physical constraints that are embedded in the network training process.
[0085] More importantly, by dynamically adjusting the weight coefficients to balance the data fitting term and the physical mechanism term, this step adaptively adjusts the constraint strength according to different states during the training phase: emphasizing the fitting of observed values when data is abundant, and emphasizing adherence to medical mechanisms when data is sparse or noisy. This dynamic balancing mechanism forces the network to simultaneously satisfy the consistency of clinical observation data and the consistency of disease evolution mechanisms during the optimization process. This ensures that the survival risk prediction output by the model is both an accurate replication of historical cases and a logical deduction of the future development trend of the disease, thereby significantly improving the model's generalization ability and reliability in complex clinical environments.
[0086] In some embodiments of this application, physical constraints are implemented by imposing penalties on solutions that deviate from the evolution of the differential equation.
[0087] Specifically, the physical constraint term is implemented by penalizing solutions that deviate from the evolutionary laws of the differential equation. This means that when the disease development trend predicted by the model (such as changes in tumor volume or fluctuations in viral load) violates the pre-set dynamic differential equation, the loss function immediately generates a huge numerical penalty, thereby strongly correcting the network parameters during backpropagation. This mechanism effectively prevents the model from fitting curves that violate medical common sense simply to cater to noisy clinical observation data (such as predicting that tumors will suddenly disappear or grow exponentially without intervention).
[0088] Through this penalty mechanism, the model is guided to a solution space that conforms to the pathophysiological mechanism, ensuring that the final prediction results are not only statistically accurate, but also highly reasonable and interpretable in terms of medical mechanism, thus solving the problem that deep learning models are prone to hallucination predictions when lacking mechanistic constraints.
[0089] In some embodiments of this application, step S400, outputting the patient's survival function within a future preset time window, specifically includes: Step S410: Based on the trained model, derive the individualized survival rate-time curve and the expected remaining survival period.
[0090] In step S410, the complex deep computational results of the model are transformed into prognostic quantitative indicators that are most intuitive and of greatest concern to clinicians and patients, realizing the mapping from abstract feature space to specific clinical decision indicators. Based on the trained PINN-Cox joint prediction model, this step can derive individualized survival rate-time curves and expected remaining survival for each patient based on their specific baseline characteristics and time-series data.
[0091] This is not merely a simple probability output, but a precise prediction that combines the patient's unique pathophysiological state and the disease's evolutionary patterns. Individualized survival curves clearly demonstrate the patient's probability of survival at different points in the future, while the expected remaining survival provides crucial time references for physicians to assess the severity of the disease and formulate treatment plans (such as whether aggressive or palliative treatment is needed), greatly enhancing the model's guiding value in practical clinical applications.
[0092] Step S420: Simultaneously invert and output the changing trends of key potential pathological states constrained by the mechanistic equation. Key potential pathological states include the in vivo viral load change trajectory or the dynamic evolution trend of tumor burden indicated by the mechanistic dynamic residual.
[0093] In step S420, the characteristics of physical information neural networks are utilized to achieve non-invasive continuous monitoring and visualization of key pathophysiological indicators. This step simultaneously inverts and outputs the changing trends of key potential pathological states constrained by mechanistic equations. These states include the in vivo viral load change trajectory or the dynamic evolution trend of tumor burden indicated by mechanistic dynamic residuals.
[0094] In clinical practice, it is often difficult to obtain continuous data on viral load or tumor burden through frequent invasive examinations. This step utilizes the dynamic laws learned by the model to extrapolate the continuous evolution trajectory of these latent variables from sparse clinical observation data. This not only verifies the reliability of the model's predictions (i.e., the prediction results are derived from reasonable pathological evolution), but also provides doctors with a "penetrating eye" into the microscopic evolution of the disease, helping them understand the specific pathological drivers behind declining survival rates.
[0095] Step S430: Generate clinical risk warning information based on the survival function and key potential pathological states.
[0096] In step S430, clinical risk warning information is generated based on the survival function and key potential pathological states. The system comprehensively analyzes the rate of decline in survival rate and abnormal fluctuations in potential pathological states (such as viral load surges or uncontrolled tumor burden), automatically identifying high-risk time windows or critical states. Once a patient's predictive indicators reach a preset risk threshold, the system immediately generates a warning, alerting medical staff that the patient may be facing an imminent risk of disease deterioration or death. This warning mechanism transforms medical intervention from "post-event remediation" to pre-event prevention, buying doctors valuable rescue window time, thereby truly achieving intelligent risk management throughout the entire process of critical illness.
[0097] In some embodiments of this application, the complete flow of the method is as follows: (1) Establish a critical care patient-level data access platform to obtain longitudinal clinical data related to patients with severe influenza or malignant tumors from the hospital information system, covering multi-dimensional information such as etiological and tumor marker quantitative results, vital signs, routine laboratory tests and clinical intervention events.
[0098] Based on this, the system will perform strict identity mapping and unit and dimension unification processing on various heterogeneous data, and set a predetermined time granularity in days or weeks to construct a standardized patient-level timeline.
[0099] By accurately mapping detection data from different frequencies and sources to discrete follow-up time points, this step can form a consistent and standardized high-dimensional time series, thus laying a solid data foundation for subsequent feature construction and model analysis.
[0100] (2) The output time series undergoes in-depth quality control and feature reconstruction. Considering the frequent occurrence of observational gaps in critical care scenarios, this step introduces an interpolation algorithm based on disease evolution mechanisms for intelligent completion and explicitly constructs missing indicator variables to retain the potential information inherent in the missing data. Simultaneously, the system extracts static features including age, gender, and baseline pathological stage, as well as dynamic evolutionary features such as the slope of biomarker fluctuations within a specific time window and the temporal changes in drug dosage.
[0101] Subsequently, these structured features are vectorized to obtain a set of input feature vectors, and key treatment interventions such as radiotherapy, chemotherapy, and ventilator use are specifically time-coded to ensure that the model can fully capture the complex details of disease progression.
[0102] (3) Construct a PINN-Cox prediction model that integrates disease mechanisms, take time variables and dynamic feature vectors as network inputs, and use the powerful nonlinear mapping capability of deep neural networks to replace the linear combination of covariates in the traditional Cox model, thereby outputting a more accurate time-dependent risk function.
[0103] During the training process, this step innovatively incorporates the evolutionary mechanisms of severe diseases, such as tumor growth kinetics equations or viral load decay equations, as physical constraints into the network. By constructing a joint loss function that includes data fitting partial likelihood loss, dynamic residual loss, and regularization term, and using the backpropagation algorithm to continuously update the network parameters, a deep joint optimization of data-driven and mechanism-constrained approaches is achieved, significantly improving the model's generalization ability and scientific interpretability.
[0104] (4) Based on the trained model, a comprehensive survival prediction is made for the target patient at the current or future time point, and key prognostic indicators such as individualized survival function curve, cumulative hazard rate and median expected life expectancy within the future preset time window are directly output.
[0105] Simultaneously, the system can also synchronously invert and output key potential state trajectories constrained by mechanistic equations, such as predicted trends in tumor burden changes. By comparing the predicted survival risk with preset clinical risk thresholds in real time, the system will automatically trigger an early warning if a sharp drop in the survival probability or a rapid deterioration in the pathological state is detected in the short term. This effectively assists the intensive care team in adjusting treatment plans in a timely manner and seizing the initiative in treatment.
[0106] In some embodiments of this application, the PINN-Cox model structure involved in this invention is as follows: (1) The main function of the input layer is to receive the time-series follow-up data after feature engineering and the corresponding timestamps. This serves as the foundational information for subsequent network processing.
[0107] (2) The deep feature extraction network adopts the Transformer architecture, which aims to learn complex nonlinear representations related to prognosis and survival from high-dimensional features, thereby replacing the linear combination method in the traditional Cox model and capturing deeper pathological feature associations.
[0108] (3) The time modeling and dynamic state layer is responsible for mapping the extracted features into potential state variables with physical meaning, such as inflammation intensity or tumor burden, and using the network to differentiate the time variable t in order to construct a dynamic description that conforms to the disease evolution law.
[0109] (4) The output layer is used to output the time-varying risk function. and the instantaneous hazard rate of the non-proportional Cox model Based on these core indicators, cumulative risk and survival probability can then be derived. This provides quantitative evidence for clinical prognosis.
[0110] Furthermore, the definition formula for the mechanism-data joint loss function is as follows: .in, This represents the Cox mean negative log-partial likelihood loss based on the approximation of the sampled risk set. The residuals of the differential equations for disease progression constructed using automatic differentiation include physical constraint terms such as residuals of the tumor Gompertz growth model or residuals of influenza virus dynamics. This is a regularization penalty term introduced to prevent the model from overfitting. and These are weight parameters.
[0111] In some embodiments of this application, to address the fusion problem of baseline static features and dynamic evolutionary features, the deep feature extraction network of this embodiment employs a feature broadcasting and concatenation strategy. Specifically, the static feature vectors such as the patient's age and gender are broadcast and copied in the time dimension, making their length consistent with the step size of the dynamic temporal features. Then, the broadcast static features and the dynamic features of each step are concatenated in the channel dimension to form an enhanced temporal input tensor. For imaging follow-up features, this embodiment does not directly input the original pixels, but first uses a pre-trained convolutional neural network (CNN) (such as ResNet-18) to extract features from DICOM format CT or MRI images, generating a low-dimensional image embedding vector, and then uses this vector as a feature to supplement the temporal data at the corresponding time step. This processing method effectively reduces the computational dimensionality while preserving key lesion information in the images.
[0112] In some embodiments of this application, to achieve a deep fusion of physical mechanisms and survival prediction, the output structure of the deep feature extraction network is specially designed. Specifically, the high-dimensional feature vector output by the Transformer encoder is fed into two parallel decoders: the first decoder is a survival risk prediction branch, which is mapped through a fully connected layer and activated by a Softmax function to output a time-dependent relative risk function. The second decoding head is the physical state deduction branch, which directly outputs latent state variables with explicit physical meaning. (e.g., tumor burden). During training, The residuals were forcibly substituted into the Gompertz differential equation for calculation, and This is then used to calculate the Cox negative log-likelihood loss. This multi-task learning architecture enables the shared Transformer underlying parameters to simultaneously capture statistical and physical laws, achieving true end-to-end joint training.
[0113] In some embodiments of this application, reference is made to Figure 2 This demonstrates the architecture of a critical illness risk prediction system based on the PINN-Cox joint model.
[0114] First, the input time variable t and multi-source temporal features X(t) are encoded and processed in high dimension using the Transformer deep feature extraction network. This deep feature extraction network serves as the shared backbone of the entire model, responsible for extracting potential feature representations that can characterize the patient's pathological evolution trend from complex clinical time-series data, providing a unified feature input foundation for subsequent data-driven and physical mechanism branches.
[0115] For the A critically ill patient, The time when the event (such as death) occurred or was deleted. For event indicator, where Indicates that an event has occurred. Indicates deletion. The set of input feature vectors.
[0116] Furthermore, in the data-driven branch, namely the Cox branch, the system constructs a relative risk feature mapping layer. In conjunction with clinical observation data Construct a deep Cox non-proportional risk architecture.
[0117] Specifically, given that the risk function of the traditional Cox model is limited by the proportionality assumption, this invention utilizes a deep neural network to construct a time-dependent relative risk function. Then the individual in time Non-proportional instantaneous hazard rate Defined as: ;in, This is the baseline risk function. Based on this, the patient's risk over time can be derived using the cumulative hazard function. Survival probability The expression is as follows: ; Furthermore, in the physical mechanism branch, i.e., the PINN branch, to avoid the model getting trapped in a black box and causing prediction distortion, this invention introduces latent physical state variables representing the essential characteristics of the disease in the intermediate layer of the network. For example, setting This refers to the tumor burden volume or viral load in the patient's body. Based on prior biomedical consensus, it is assumed that the natural evolution of this state and the response to therapeutic intervention follow a specific ordinary differential equation (ODE), such as the expression of the classic Gompertz growth arrest model, as follows: ; in, Endogenous growth rate, To the maximum environmental carrying capacity, Based on the dynamic characteristics of patients The treatment clearance rate network is determined by factors such as drug dosage and immune indicators. The neural network directly calculates the state output versus time using automatic differentiation techniques. derivative Thus, physical mechanism residuals are constructed.
[0118] Furthermore, during the training phase, at the joint optimization layer, this invention abandons the single cross-entropy or likelihood loss and innovatively proposes a joint loss function. ; Among them, the data fitting term The average negative log-partial likelihood function of the non-proportional Cox model satisfies the following expression: In the formula, In time A group of patients who are still at risk of dying. This represents the number of samples in which the event occurred.
[0119] Physical mechanism constraint terms This is achieved by penalizing the neural network output of solutions that do not conform to the evolution law of the differential equation, satisfying the following expression: ; In the formula, The number of configuration points. This is achieved by introducing... The weights are dynamically adjusted to balance data-driven and mechanism-driven approaches. This architecture forces the predictive model to ensure that its internally extracted risk representations conform to the true pathodynamics of severe illness when fitting the survival distribution, greatly enhancing the generalization robustness under very small sample sizes.
[0120] Furthermore, This is a regularization penalty term introduced to prevent model overfitting. This regularization term is typically constructed using the common L1 or L2 paradigm. By imposing additional penalty costs on the network weight parameters, it effectively limits the disordered growth of model complexity. This mechanism not only helps smooth the optimization surface of the loss function but also encourages the model to maintain parameter sparsity or numerical stability while learning data features. This avoids the model over-memorizing noisy data from the training set, ensuring that it can still output robust and reliable prediction results when faced with unknown clinical samples.
[0121] In some embodiments of this application, for Gompertz differential equations Item in To address the numerical instability issue caused by values approaching zero, this embodiment introduces a numerical smoothing constraint mechanism at the output of the physical state derivation branch. Specifically, before calculating the physical residual loss, the potential state variables output by the network are... conduct Processing, among which The preset minimum smoothing factor has a value of [value missing]. Alternatively, the Softplus activation function can be used as the activation function for the output layer to enforce this. The output value is always greater than 0. This mechanism effectively prevents gradient explosion or NaN values, ensuring the numerical stability of the PINN-Cox model during backpropagation.
[0122] In some embodiments of this application, the triggering logic for multi-level alarms is based on preset mathematical judgment rules. Specifically, the system monitors the patient's survival function in real time. Its derivative and rate of change trigger the corresponding level of alert when any of the following conditions are met: Yellow alert indicates survival probability. Below the first threshold (e.g., 0.5) and the slope of the risk function The trend is upward; a red alert is triggered when the survival probability S(t) is below the second threshold (e.g., 0.3) or there are potential pathological state variables. growth rate Exceeding the medical threshold. Furthermore, the interpretability attribution assistance function specifically employs the SHAP algorithm to calculate the contribution of each clinical feature to the current risk prediction, and highlights the top-3 feature variables with the highest contribution to risk spikes in the form of a heatmap on the terminal interface, thereby transforming the model's decision-making basis into understandable clinical logic.
[0123] Secondly, this application provides a critical illness risk prediction system based on the PINN-Cox joint model, characterized in that it includes a server for data processing and model inference; The server includes a processor, memory, and a communication interface. The memory stores the computer program, the communication interface is used to connect to the hospital information system to obtain multi-source time-series data of patients, and the processor is used to execute the computer program to implement the critical illness risk prediction method based on the PINN-Cox joint model as described above. The server is also configured with: (1) Mechanism-data dual-drive inference engine, used to load the trained PINN-Cox joint prediction model, perform forward inference on the input patient time-series features, and output individualized survival function and potential pathological state trajectory; (2) Risk warning and visualization module, which is used to dynamically generate multi-level risk warning signals based on the decline slope of the survival function and the degree of deterioration of potential pathological conditions, and render individual survival curves and risk trend diagrams on the connected display terminal.
[0124] Furthermore, its multimodal communication interface is based on standard medical data protocols such as HL7 and FHIR, enabling seamless integration with hospital electronic medical record systems, laboratory information systems, medical imaging systems, and bedside intensive care equipment. Simultaneously, the real-time dynamic capture subunit captures unstructured and structured dynamic data such as vital sign fluctuations, test results, and medical order execution records of target patients in real time according to a preset acquisition frequency, ensuring the comprehensiveness and timeliness of data sources.
[0125] In some embodiments of this application, the risk warning and visualization module is dedicated to transforming complex calculation results into intuitive clinical decision-making basis. The individual survival curve rendering engine dynamically draws the patient's exclusive survival rate-time evolution curve on the UI interface of the doctor's terminal. This is equivalent to an individualized Kaplan-Meier prediction curve, and intuitively marks the median expected remaining life and confidence interval.
[0126] In addition, a multi-dimensional triggering mechanism is set up. When the system detects that the predicted survival probability falls below the safety threshold or the mechanistic dynamic residual indicates that the disease has entered an out-of-control period, it will push yellow, orange, and red multi-level alarms to the responsible physician's computer desktop or mobile terminal according to the severity of the crisis. When a risk alarm is triggered, it will not only report good news and bad news, but also use feature attribution algorithms to highlight the core driving variables that caused the risk to surge, thereby providing doctors with clear logical support and closed-loop management for timely intervention and adjustment of treatment plans.
[0127] Furthermore, embodiments of this application provide an electronic device, including a processor and a memory. The memory is used to store a computer program. The processor is used to execute the computer program stored in the memory to implement the aforementioned method for predicting the risk of severe illness based on the PINN-Cox joint model.
[0128] Furthermore, embodiments of this application provide a computer-readable storage medium storing computer program instructions thereon, which, when executed by a processor, implement the aforementioned method for predicting the risk of severe illness based on the PINN-Cox joint model.
[0129] In summary, the critical illness risk prediction method and related equipment based on the PINN-Cox joint model provided in this application have the following technical effects.
[0130] First, it significantly improves the accuracy and robustness of severe illness risk prediction. This application breaks through the excessive reliance of traditional pure data-driven models on large-scale labeled data. By introducing a physical mechanism term based on automatic differentiation into the loss function, it embeds medical prior knowledge such as tumor growth or virus dynamics as strong constraints into the neural network. This mechanism effectively solves the overfitting problem caused by sparse and noisy clinical data, ensuring that the model strictly follows the objective biological laws of disease evolution while fitting clinical observation data. Thus, it can still output highly reliable prediction results in small sample or complex noisy environments.
[0131] Secondly, this application achieves white-box monitoring and interpretable prediction of the evolution of severe diseases. Unlike traditional black-box models that only output a single risk score, this application utilizes the inversion capability of physical information neural networks to simultaneously deduce and output key potential pathological states constrained by mechanistic equations, such as the trajectory of viral load changes or the dynamic evolution trend of tumor burden. This not only allows doctors to intuitively see the specific pathological drivers leading to changes in survival rates but also provides clinicians with a continuous, non-invasive means of monitoring pathological states, greatly enhancing the model's interpretability and clinical reliability.
[0132] Finally, a closed-loop diagnosis and treatment support system was constructed, encompassing real-time data perception and intelligent risk warning. Through the collaborative work of the physical computing device layer, data interaction layer, and application service layer, this application achieves seamless integration with hospital information systems and bedside monitoring equipment, enabling real-time capture of multimodal vital signs and laboratory data. Combined with the efficient computation of the mechanism-data dual-driven inference engine, the system can instantly generate individualized survival curves and push multi-level clinical risk warnings, helping medical staff seize valuable intervention windows before the condition deteriorates, thereby significantly improving the intelligence level of intensive care and the success rate of patient treatment.
[0133] It should be noted that in all specific embodiments of this application, all data processing activities related to user identity or personal characteristics, such as user information, user behavior data, historical data, and location information, will be conducted in accordance with the principles of legality, legitimacy, and necessity. All data collection, use, storage, and processing will be subject to compliance with applicable national and regional laws, regulations, and industry standards, and informed consent from users will be obtained in a clear and explicit manner before processing. For the processing of sensitive personal information, separate consent from users will be obtained through prominent means such as pop-up prompts and independent confirmation pages. If any processing conflicts with laws and regulations, the laws and regulations will prevail, and necessary data processing will only be carried out within the scope permitted by laws and regulations, ensuring that all data-based applications, analyses, and technical implementations are conducted within the scope permitted by laws and regulations.
[0134] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this application are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and sub-operations described as part of a larger operation are executed independently.
[0135] Furthermore, although this application is described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding this application. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of ordinary skill of an engineer. Therefore, those skilled in the art can implement the application set forth in the claims using ordinary skill. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of this application, which is determined by the full scope of the appended claims and their equivalents.
[0136] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several programs to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0137] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequential list of executable programs for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, a program execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can retrieve and execute a program from or in conjunction with such a program execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can mean any means that can contain, store, communicate, propagate, or transmit a program for use by or in conjunction with a program execution system, apparatus, or device.
[0138] More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Additionally, computer-readable media can even be paper or other suitable media on which programs can be printed, for example, by optically scanning the paper or other media, then editing, interpreting, or, if necessary, processing it in a suitable manner to obtain the program electronically, and then storing it in computer memory.
[0139] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable program execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0140] In the foregoing description of this specification, the reference to terms such as "one embodiment / implementation," "another embodiment / implementation," or "certain embodiments / implementations," etc., indicates that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in an embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0141] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.
[0142] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.
Claims
1. A method for predicting the risk of severe illness based on the PINN-Cox joint model, characterized in that, Includes the following steps: Acquire multi-source time-series clinical data of patients with target severe diseases, perform structured processing and time alignment, and construct an individual survival analysis dataset; The acquisition of multi-source time-series clinical data from patients with target critical illnesses, followed by structured processing and time alignment, specifically includes: acquiring longitudinal clinical data from hospital electronic medical record systems or intensive care databases, including etiological and tumor marker indicators, vital signs and physiological and biochemical indicators, imaging follow-up characteristics, and clinical treatment events; performing de-identification, outlier cleaning, unit standardization, and missing value processing on the acquired longitudinal clinical data; and constructing a patient-level timeline with a predetermined time granularity, mapping observation records of different frequencies to corresponding time points to achieve time alignment. A PINN-Cox joint prediction model is constructed, which inputs time variables and dynamic feature vectors into a deep neural network and outputs a time-dependent relative risk function through nonlinear mapping to characterize the instantaneous hazard rate under the non-proportional Cox model. The construction of the PINN-Cox joint prediction model specifically includes: constructing an input layer to receive time-series follow-up data and corresponding timestamps after feature engineering; constructing a deep feature extraction network using a Transformer architecture to learn complex nonlinear representations related to prognosis and survival from high-dimensional features; constructing a time modeling and dynamic state layer to calculate the derivative of the neural network output with respect to time using automatic differentiation techniques to deduce physically meaningful latent state variables, including inflammation intensity or tumor burden; and constructing an output layer to output a time-varying risk function and the instantaneous hazard rate of the non-proportional Cox model. Define a mechanism-data joint loss function, and use the multi-source time-series clinical data to jointly constrain the training of the PINN-Cox joint prediction model; the mechanism-data joint loss function includes a negative log-partial likelihood loss term based on the non-proportional Cox model, and a physical residual loss term based on the differential equation of the evolution mechanism of severe diseases; The mechanism-data joint loss function is specifically defined as follows: the total loss equals the weighted sum of the data fitting term loss, the physical mechanism term loss, and the regularization term loss; wherein, the data fitting term loss is the Cox mean negative log-partial likelihood loss based on the sampling risk set approximation; the physical mechanism term loss is the residual of the differential equation for disease progression constructed using automatic differentiation, specifically the residual of the dynamic model characterizing the pathological evolution of the target organ, including the residual of the tumor Gompertz growth model or the residual of the influenza virus dynamics; and the regularization term loss is an L1 or L2 norm penalty term. Input the time series data of the patient to be predicted into the trained model, and output the patient's survival function, cumulative risk function and median expected life expectancy within a future preset time window.
2. The method according to claim 1, characterized in that, The missing value processing specifically involves: using a combination of nearest-neighbor observation imputation and interpolation estimation based on disease dynamics to complete the missing values and generate missing indicator variables; the method also includes extracting key features characterizing disease evolution and mortality risk, the key features including baseline static features and dynamic evolution features, wherein the dynamic evolution features include indicator level, rate of change, fluctuation amplitude and stage mutation information.
3. The method according to claim 1, characterized in that, The joint constraint training of the PINN-Cox joint prediction model using the multi-source time-series clinical data specifically includes: The gradient of the physical residual loss term is calculated using automatic differentiation techniques, and the network parameters are updated synchronously during backpropagation. The evolutionary mechanism of severe diseases is embedded as a physical constraint in the network training process. By dynamically adjusting the weight coefficients to balance the data fitting term and the physical mechanism term, the network is forced to satisfy the consistency of clinical observation data and the consistency of disease evolution mechanism during the optimization process. The physical constraint term is implemented by imposing penalties on solutions that deviate from the evolution law of the differential equation.
4. The method according to claim 1, characterized in that, The output of the patient's survival function within a future preset time window specifically includes: Based on the trained model, individualized survival rate-time curves and expected remaining survival are derived. The changes in key potential pathological states constrained by mechanistic equations are simultaneously inverted and output. These key potential pathological states include the in vivo viral load change trajectory or the dynamic evolution trend of tumor burden indicated by mechanistic dynamic residuals. Based on the survival function and the key potential pathological states, clinical risk warning information is generated.
5. A critical illness risk prediction system based on the PINN-Cox joint model, characterized in that, This includes servers used for data processing and model inference; The server includes a processor, a memory, and a communication interface. The memory stores a computer program, and the communication interface is used to connect to a hospital information system to obtain multi-source time-series data of patients. The processor is used to execute the computer program to implement the critical illness risk prediction method based on the PINN-Cox joint model as described in any one of claims 1 to 4. The server is also configured with: The mechanism-data dual-driven inference engine is used to load the trained PINN-Cox joint prediction model, perform forward inference on the input patient time-series features, and output individualized survival functions and potential pathological state trajectories. The risk warning and visualization module is used to dynamically generate multi-level risk warning signals based on the declining slope of the survival function and the degree of deterioration of potential pathological conditions, and to render individual survival curves and risk trend graphs on the connected display terminal.
6. An electronic device, characterized in that, Including processor and memory; The memory is used to store computer programs; The processor is used to execute the computer program stored in the memory to implement the critical illness risk prediction method based on the PINN-Cox joint model as described in any one of claims 1 to 4.
7. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the critical illness risk prediction method based on the PINN-Cox joint model as described in any one of claims 1 to 4.
Citation Information
Patent Citations
Transistor high-power electromagnetic thermal model optimization system based on deep learning
CN120409229A
Methods, systems, and computer readable media for causal training of physics-informed neural networks
US20250077875A1
