A program, apparatus, and method for predicting postoperative outcomes based on human time-series biological information.

Abstract

This invention provides a postoperative complication prediction program and system, as well as a method for generating a postoperative complication prediction model and a trained machine learning model, which can improve patient treatment outcomes by predicting the risk of complications in patients after surgery and enabling appropriate management. [Solution] A new postoperative complication prediction program specifically for the surgical field, which integrates and analyzes diverse clinical information, including time-series biological information obtained from wearable devices, etc., to predict the risk of complications occurring within, for example, 90 days after surgery, both early and over time.

Description

Technical Field

[0005]

[0001] The present disclosure relates to a program, apparatus, and method for predicting the risk of postoperative complications using patient biometric information and the like.

Background Art

[0002] Postoperative complications cause increased mortality, reoperations, longer hospital stays, and increase medical costs and the burden on patients. If the risk can be predicted early, appropriate management can be implemented to improve the outcome.

[0003] Conventionally, several studies on factors for predicting the risk of postoperative complications have been reported. For example, a research group at the University of Florida in the United States has developed an AI platform for predicting postoperative complications using clinical data extracted from electronic medical records and machine learning algorithms (Non-Patent Document 1). However, conventional predictions were based on data at a single point in time.

[0004] In addition, the usefulness of electronic medical record data and vital signs has also been shown in risk prediction models for disease deterioration (Non-Patent Document 2). Also in Japan, the development of program medical devices (SaMD) utilizing AI technology has progressed, and image diagnosis support software, atrial fibrillation prediction programs, etc. have been approved as medical devices. However, these technologies were mainly specialized in diagnostic support and the emergency and intensive care fields (Non-Patent Document 3).

Prior Art Documents

Non-Patent Documents

[0005]

Non-Patent Document 1

Non-Patent Document 2

[0006] Predicting and detecting postoperative complications earlier can improve patient outcomes. There is a need for new methods that can predict postoperative complications more accurately than conventional methods. [Means for solving the problem]

[0007] The invention described herein has been made in view of this problem, and its purpose is to improve patient treatment outcomes by predicting the risk of complications in patients after surgery and enabling appropriate management. As a means to solve this problem, this disclosure provides a novel postoperative complication prediction program specifically for the surgical field, as well as related methods, apparatus, and recording media, which comprehensively analyzes diverse clinical information, including time-series biological information obtained from wearable devices, etc., to predict the risk of complications occurring, for example, within 90 days after surgery, both early and over time. Many existing methods rely on data at a single point in time, and have not adequately utilized continuous or time-series data, which the inventors emphasize. On the other hand, the advantage of using time-series data is that not only does the amount of information increase, but changes over time and differences in individual patient responses are also taken into account, making it useful for predicting disease states.

[0008] This disclosure provides, more specifically, the following embodiments. [Embodiment 1] A step to obtain time-series biological information from before surgery to a predetermined period after surgery in patients who have undergone surgical procedures, A model processing step that outputs information regarding postoperative complications based on the acquired time-series biological information. Have the computer run it, A postoperative complication prediction program characterized by its ability to predict complications that will occur within 90 days after surgery. [Embodiment 2] The postoperative complication prediction program according to Embodiment 1, characterized in that the model processing step further uses, in addition to the time-series biological information, at least one piece of information selected from the group consisting of patient physical information, comorbidity information, treatment target disease information, and surgical information, which are data from a single point in time, as input, and outputs information regarding the complications. [Embodiment 3] The postoperative complication prediction program according to Embodiment 1 or 2, characterized in that the model processing step further uses surgical images or videos as input. [Embodiment 4] The first predictive model outputs the risk of developing all postoperative complications. A second predictive model outputs the complications that are predicted to occur frequently in the target patient. A third predictive model that outputs the risk of developing complications that frequently occur in specific surgeries. A fourth predictive model that outputs the risk of death, A postoperative complication prediction program according to any one of embodiments 1 to 3, characterized in that it outputs information regarding the complication using one, two, three, or four predictive models selected from the group consisting of the following. [Embodiment 5] The postoperative complication prediction program according to any one of embodiments 1 to 4, characterized in that the model processing step outputs a predetermined number of high-risk complication risks using at least two models. [Embodiment 6] A postoperative complication prediction program according to any one of Embodiments 1 to 5, characterized in that, as information regarding the aforementioned complication, it issues a warning when the risk of developing the complication exceeds a predetermined threshold. [Embodiment 7] A postoperative complication prediction program according to any one of embodiments 1 to 6, characterized in that, as information regarding the aforementioned complications, it outputs information suggesting 1 to 50 recommended tests or treatments according to the risk of onset calculated for the purpose of reducing the risk of developing or preventing the worsening of postoperative complications. [Embodiment 8] A postoperative complication prediction program according to any one of Embodiments 1 to 7, further comprising a function that uses generative artificial intelligence to provide input support for the input information, to answer questions regarding the input or output information, or to suggest optimal input parameters. [Embodiment 9] A postoperative complication prediction program according to any one of embodiments 1 to 8, characterized in that predicted values ​​are calculated over time, reflecting the input of data over time after surgery. [Embodiment 10] A postoperative complication prediction program according to any one of embodiments 1 to 9, characterized in that it outputs complication risk and / or alerts in real time. [Embodiment 11] A postoperative complication prediction program according to any one of Embodiments 1 to 10, characterized by proposing factors, change targets, and methods that can improve postoperative outcomes before surgery. [Embodiment 12] A postoperative complication prediction program according to any one of embodiments 1 to 11, characterized by having a feedback function that collects data, uses that data to identify new or highly accurate factors or methods useful for predicting complications, and improves the accuracy of complication prediction. [Embodiment 13] A postoperative complication prediction program according to any one of embodiments 1 to 12, wherein a display device connected to a computer displays, as a user interface, a first area for inputting single-point-of-time data, a second area for inputting time-series data, a third area for accepting operations to perform analysis of a postoperative complication prediction model, and a fourth area for outputting the results of the model analysis. [Embodiment 14] The second area has a manual data input function, a data linking function with a measuring device, and a function of displaying the temporal change of the input time-series data as a chart. The fourth area has an output form according to the risk level of postoperative complications and includes a warning function of presenting a warning when the risk exceeds a predetermined threshold. The postoperative complication prediction program according to Embodiment 13. [Embodiment 15] A step of acquiring time-series biological information of a patient who has undergone surgery from before surgery to a predetermined period after surgery; A step of inputting the acquired time-series biological information into a learned artificial intelligence model and acquiring information on postoperative complications output by the model A method for predicting postoperative complications that occur within 90 days after surgery, including. [Embodiment 16] An acquisition unit that acquires time-series biological information of a patient who has undergone surgery from before surgery to a predetermined period after surgery; A processing unit that inputs the acquired time-series biological information into a learned artificial intelligence model and acquires information on postoperative complications output by the model A prediction device for postoperative complications that occur within 90 days after surgery, including. [Embodiment 17] A computer-readable recording medium recording the postoperative complication prediction program according to any one of Embodiments 1 to 14. [Embodiment 18] At least one measuring device that acquires time-series biological information of a patient who has undergone surgery from before surgery to a predetermined period after surgery; A processing device configured to receive the time-series biological information from the measuring device and execute a model processing step of outputting information on postoperative complications that occur within 90 days after surgery based on the time-series biological information A postoperative complication prediction system, characterized by comprising. [Embodiment 19] A step of acquiring, for a plurality of past patient cases, time-series biological information from before surgery to a predetermined period after surgery and performance data regarding the occurrence or non-occurrence of complications within 90 days after surgery; A step of training a machine learning model for predicting postoperative complications, using the time-series biological information as input and the performance data as correct labels A method for generating a postoperative complication prediction model, characterized by including the above [Embodiment 20] A trained machine learning model generated by the method described in Embodiment 19

Effect of the Invention

[0009] According to the present disclosure, it becomes possible to predict the onset risk of postoperative complications early and over time. Thereby, preventive treatment intervention according to the risk is realized, contributing to the improvement of the patient's prognosis. In addition, by cooperating with a wearable device, continuous and automatic monitoring of biological information becomes possible, and the physical burden on the patient and the workload of medical staff can also be reduced

Brief Description of the Drawings

[0010] [Figure 1] It is a diagram showing an example of a user interface of a program according to an embodiment of the present invention [Figure 2] It is a diagram showing another example of a user interface of a program according to an embodiment of the present invention [Figure 3] It is a diagram showing another example of a user interface of a program according to an embodiment of the present invention [Figure 4] It is a diagram showing another example of a user interface of a program according to an embodiment of the present invention [Figure 5] It is a diagram showing another example of a user interface of a program according to an embodiment of the present invention [Figure 6] It is a diagram showing another example of a user interface of a program according to an embodiment of the present invention [Figure 7] It is a diagram showing another example of a user interface of a program according to an embodiment of the present invention [Figure 8] It is a diagram showing another example of a user interface of a program according to an embodiment of the present invention [Figure 9] This figure shows another example of a user interface for a program relating to one embodiment of the present invention. [Figure 10] This figure shows another example of a user interface for a program relating to one embodiment of the present invention. [Figure 11] This is a flowchart of a program relating to one embodiment of the present invention. [Figure 12] This is a functional block diagram of a device relating to one embodiment of the present invention. [Figure 13] This diagram illustrates the flow of information in a program related to one embodiment of the present invention. [Figure 14] This diagram illustrates an example of model combinations and feedback functions. [Modes for carrying out the invention]

[0011] The embodiments for carrying out the present invention will be described below. The postoperative complication prediction program according to this disclosure (hereinafter also referred to as "the program according to this disclosure") is provided as a program that runs on a computer such as a desktop PC, notebook PC, tablet, or smartphone. The descriptions in this disclosure also apply to methods, apparatus, and recording media related to the program according to this disclosure.

[0012] (program) One aspect of this disclosure relates to a postoperative complication prediction program. More specifically, the program relating to this disclosure may be a postoperative complication prediction program characterized by causing a computer to perform the steps of acquiring time-series biological information of a patient who has undergone surgery from before surgery to a predetermined period after surgery, and a model processing step of outputting information regarding postoperative complications based on the acquired time-series biological information.

[0013] (Forecast period) The program relating to this disclosure can predict, for example, complications that occur within 90 days, 60 days, 30 days, 20 days, 10 days, 7 days, 5 days, 4 days, 3 days, 2 days, or 1 day postoperatively. Preferably, the program relating to this disclosure predicts complications that occur within 30 days postoperatively.

[0014] (Postoperative complications) Postoperative complications can be, but are not limited to, predictions of all postoperative complications, intra-abdominal infection complications, and / or pneumonia. Types of complications include abscess, intra-abdominal infection complications, intra-abdominal abscess, ascites, empyema, pleural effusion, pneumothorax, tracheal stump fistula, anastomotic leakage, anastomotic stricture, pancreatic fistula, pancreatitis, hemorrhage, pneumonia, atelectasis, respiratory failure, acute respiratory distress syndrome, dysphagia, pyelonephritis, cystitis, prostatitis, urinary tract infection, cholecystitis, cholangitis, biliary tract infection, catheter-related infection, sepsis, bacteremia, wound infection, cellulitis, vasculitis, osteomyelitis, arthritis, encephalitis, meningitis, prosthetic infection, shock, and bowel obstruction. These conditions can be wide-ranging, including intussusception, intestinal necrosis, enteritis, gastrointestinal perforation, delayed gastric emptying, reflux esophagitis, dumping syndrome, lymph leakage, myocardial infarction, angina pectoris, heart failure, arrhythmia, cerebral infarction, cerebral hemorrhage, subarachnoid hemorrhage, seizures, epilepsy, hydrocephalus, liver damage, kidney damage, thromboembolism, fat embolism, air embolism, arterial dissection, edema, hernia, hypoglycemia, electrolyte abnormalities, allergies, drug-induced organ damage, organ injury during surgery, cardiopulmonary arrest, and death.

[0015] (surgery) Surgical procedures may include, but are not limited to, thoracic surgery, abdominal surgery, thoracoabdominal surgery, head and neck surgery, digestive system surgery, digestive system surgery, malignant tumor surgery of the digestive system, upper gastrointestinal surgery, stomach surgery, orthopedic surgery, plastic surgery, cosmetic surgery, obstetric surgery, gynecological surgery, breast surgery, urological surgery, head and neck surgery, otolaryngological surgery, oral surgery, dental surgery, cardiac surgery, vascular surgery, respiratory surgery, neurosurgery, dermatological surgery, ophthalmic surgery, pediatric surgery, endocrine surgery, transplant surgery, trauma surgery, and emergency surgery. Organs targeted by surgery may also include, but are not limited to, the oral cavity, pharynx, larynx, esophagus, stomach, duodenum, small intestine, large intestine, anus, liver, gallbladder, bile duct, pancreas, spleen, lungs, brain, spinal cord, heart, mediastinum, uterus, ovaries, mammary glands, adrenal glands, kidneys, bladder, ureters, prostate, bone, muscle, skin, blood vessels, and lymph nodes. Digestive surgery may include surgeries on the pharynx, larynx, esophagus, stomach, duodenum, small intestine, large intestine, anus, liver, gallbladder, bile duct, pancreas, and spleen. Compared to surgeries for benign diseases, surgeries for malignant tumors are more difficult and have a higher incidence of postoperative complications. Different predictive models may be used for surgeries for benign diseases and malignant tumors. Surgeries for malignant tumors include, for example, cancer, sarcoma, endocrine tumors, GIST, mesenchymal tumors, neurological tumors, germ cell tumors, embryonic tumors, malignant melanoma, lymphoma, hematological malignancies, hematopoietic malignancies, and tumors of unknown primary origin. Surgery for benign diseases includes, for example, gastrointestinal perforation, gastrointestinal stricture, biliary stricture, foreign body in the gastrointestinal tract, gastrointestinal bleeding, splenic bleeding, hepatic bleeding, organ bleeding, intestinal obstruction, colostomy, gastrostomy, appendicitis, diverticulitis, cholecystitis, peritonitis, bile duct stones, pancreatitis, hernia, reflux esophagitis, hemorrhoids, anal fistula, anal diseases, rectal prolapse, organ prolapse, intra-abdominal abscess, abscess, abdominal trauma, chest trauma, pneumothorax, hemothorax, empyema, pleural effusion, lung cyst, pectus excavatum, airway foreign body, myocardial infarction, angina pectoris, valvular heart disease, congenital heart disease, arrhythmia, aortic dissection, aortic aneurysm, arterial stenosis, arterial thrombosis, and venous thrombosis. These surgeries are performed for conditions such as varicose veins, hydrocephalus, cerebral hemorrhage, subdural hematoma, epidural hematoma, cerebral aneurysm, arteriovenous malformation, fractures, spinal stenosis, herniated discs, osteoarthritis, necrotizing fasciitis, benign prostatic hyperplasia, urinary tract stones, testicular torsion, hydrocele, undescended testicles, urinary incontinence, urethral stricture, hydronephrosis, nephrostomy, bladder prolapse, cataracts, sinusitis, tonsillar enlargement, tonsillitis, otosclerosis, otitis media, sleep apnea syndrome, epistaxis, temporomandibular joint disorder, salivary gland diseases, cesarean section, childbirth, miscarriage, uterine prolapse, ovarian hemorrhage, ovarian torsion, burns, and benign tumors of the aforementioned organs.

[0016] (Period for acquiring biometric information) The period from before surgery to a predetermined period after surgery for acquiring time-series biological information may be, for example, from 1 minute before surgery to 30 days before surgery, such as 10 days, 9 days, 8 days, 7 days, 6 days, 5 days, 4 days, 3 days, 2 days, 1 day before, or 1 minute before surgery, to the day of surgery or from 1 to 30 days after surgery, such as 1 day, 2 days, 3 days, 4 days, 5 days, 6 days, 7 days, 8 days, 9 days, 10 days, 15 days, 20 days, 25 days, or 30 days after surgery, but is not limited to this period. Preferably, information from within 7 days before surgery is used. Generally, postoperative complications occur within 30 days after surgery, and most occur within 10 days.

[0017] This information can be entered manually, uploaded as a file, or linked to an electronic medical record system. The time-series graphs may display vital data, other clinical data, blood biochemistry test results, and patient symptom information.

[0018] (Overview of information analysis and output) In one embodiment of this disclosure, various input data are processed by one or more pre-constructed analytical models. These models can be constructed using a single method or a combination of methods depending on the specific purpose. For example, statistical methods, machine learning, mathematical optimization, or hybrid approaches thereof may be used. For static data (e.g., patient physical information, comorbidities), statistical and machine learning methods such as logistic regression analysis, generalized linear models, support vector machines (SVMs), and decision trees can be applied. On the other hand, for time-series data such as vital signs, and multimedia data such as medical images and surgical videos, more advanced machine learning algorithms such as deep learning or generative AI can be used to capture their structural and temporal characteristics.

[0019] (Model building: Data preparation) The aforementioned model is constructed, for example, by the following procedure. First, data is prepared for a large number of past cases, including single-point-of-time data (patient information, disease information, surgical information, etc.) for pre-operative, intra-operative, and post-operative periods, as well as time-series biological information (vital data, clinical data, clinical test values, medical images, symptoms, etc.) for pre-operative, intra-operative, and post-operative periods, and data containing actual events that occurred (e.g., presence, type, onset time, severity, and treatment performed) as ground truth labels. This data can serve as training data for learning the model. Furthermore, unsupervised learning methods such as clustering and dimensionality reduction can be applied using data without ground truth labels, semi-supervised learning that utilizes a small amount of labeled data and a large amount of unlabeled data, and self-supervised learning that generates pseudo-labels from the data itself for learning. Preprocessing such as imputation of missing values, removal of outliers, and normalization / standardization of data is performed as appropriate during the data preparation stage. Furthermore, generative models such as variational autoencoders (VAEs), generative adversarial networks (GANs), and diffusion models can be used to artificially generate data on rare cases and utilize it for data augmentation to correct imbalances in training data.

[0020] (Model construction: Feature extraction and final output) Next, features useful for prediction are extracted or generated using arbitrary items from the prepared data. For example, from time series data, recurrent neural networks (RNNs), their derivatives such as LSTM (Long Short-Term Memory), GRU (Gated Recurrent Unit), or Transformer architectures with attention mechanisms, or one-dimensional convolutional neural networks (1D-CNNs) are used to capture their temporal dependencies and patterns. From image and video data, convolutional neural networks (CNNs) (e.g., ResNet, EfficientNet, etc.) or Vision Transformers (ViT) are used to extract spatial features. For example, a 2D-CNN model based on ResNet18 could be constructed as a Deep Learning model using pre-operative CT image data. Alternatively, unsupervised learning methods such as autoencoders or principal component analysis (PCA) may be used to obtain low-dimensional features (latent variables) that represent the essential structure of the data. Furthermore, for graph structure data that represents relationships such as those between patients, molecular structures, and gene networks, it is also possible to extract features using graph neural networks (GNNs).

[0021] Then, the extracted and generated features are arbitrarily combined with static data and / or time-series data and input into a model to produce the final output. Examples of such models include, but are not limited to, neural networks, gradient boosting decision trees (e.g., XGBoost, LightGBM), and random forests. Transfer learning, which utilizes knowledge from a model pre-trained on another task, and fine-tuning, which adapts that model to a specific task, can also be used.

[0022] Furthermore, for problems requiring sequential decision-making, such as selecting a treatment plan, it is possible to apply reinforcement learning to learn the optimal course of action.

[0023] (Model building: training and optimization) Model training is performed by optimizing a defined objective function (such as a loss function) to output the probability of a specific event occurring or the predicted value of a specific indicator. The specific optimization method varies depending on the model. For example, when using a neural network as the model, parameters are updated using stochastic gradient descent (SGD) or its derivatives (e.g., Adam, RMSprop) based on the gradient calculated by backpropagation. When using a gradient boosting decision tree, the model is built by sequentially adding new learners using the gradient of the loss function.

[0024] Furthermore, ensemble learning (e.g., bagging, boosting, stacking) can be used to integrate predictions from multiple models to improve overall performance. Hyperparameter tuning may also be performed using techniques such as grid search, random search, and Bayesian optimization to maximize model performance.

[0025] (Model evaluation) After building the model, its generalization performance is evaluated using an independent validation dataset that was not used for training. Depending on the purpose, evaluation metrics such as Area Under the Curve (AUC), Precision, Recall, F-score, Accuracy, and Mean Squared Error (MSE) can be used individually or in combination. To improve the robustness of the evaluation, it is desirable to employ cross-validation, which involves dividing the data into multiple groups and repeating training and validation. Furthermore, to understand the basis of the model's predictions and ensure its reliability, methods such as SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) can be used to visualize and analyze which input variables contributed to the predictions.

[0026] (Input variable) The data used as input variables in the aforementioned model is not limited to time-series biological information, but can also broadly include data from a single point in time and qualitative information. Specifically, it is possible to arbitrarily combine and use patient physical information (height, weight, BMI, etc.), information on comorbidities, medical history, surgical history, medication information (diabetes, smoking history, etc.), information on diseases being treated (disease name, progression, etc.), surgical information (surgical procedure, surgery time, etc.), clinical test values, medication information, genetic information, and even lifestyle data obtained from wearable devices.

[0027] The model processing step may output information about complications by further using at least one piece of information selected from a group consisting of time-series biological information such as vital data, clinical data, symptom data, blood test data, and laboratory information, or single-point-time data such as patient background information, patient physical information, comorbidity information, medical history, medication information, target disease information, and surgical information as input. The model processing step may further output information about complications by using surgical images or videos as input.

[0028] Figure 14 is a schematic diagram showing an example of the construction and feedback function of the complication prediction model related to this disclosure. The complication prediction system related to this disclosure may include a mechanism for sequentially improving model accuracy by analyzing patient information, surgical information, biological data, examination data, image / video data, etc., in a multi-layered manner, and may include the following steps or components. (1) Conventional predictive model construction process This process involves constructing a conventional complication prediction model using single-point-of-time data such as patient information and surgical information. (2) Intermittent time series data model construction process This process involves constructing a novel complication prediction model, which has not been previously reported, using intermittently acquired time-series data such as blood test data and radiographic image data. (3) Continuous time series data model construction process This process involves constructing an original complication prediction model using continuous time-series data such as wearable device data, electrocardiogram data, and electroencephalogram data, and employing machine learning and other techniques. (4) Process of constructing a composite model A step of constructing a new complication prediction model by combining the multiple prediction models obtained by (1) to (3) above. (5) Process for identifying useful factors and methods In the model construction process described in (2) to (4) above, statistical and machine learning methods such as logistic regression analysis, generalized linear models, support vector machines, decision trees, and deep learning are used to select factors and combinations that can predict complications with high accuracy from conventionally reported single-time period data. In addition, novel factors and combinations that contribute to the accuracy of complication prediction are extracted from time-series data such as wearable devices and surgical images and videos that have not been used for complication prediction until now. (6) Process to add feedback function A step comprising incorporating the factors or methods identified in (5) above into the complication prediction model and providing a feedback function to update the model structure. (7) Process for building a medical device conformity model This process involves constructing a complication prediction model with adjusted input and output parameters so that it can be used as a programmable medical device (SaMD). The model is composed of any one of the above (1) to (6), or a combination thereof. (8) Data collection process A process of collecting a large amount of patient data using the aforementioned SaMD. (9) Identification of additional useful factors and methods A step of analyzing the data collected by the aforementioned SaMD and identifying new factors or methods that do not exist in the past, or that contribute to improved accuracy and are useful for predicting complications, in accordance with step (5). (10) Refeedback process A step of incorporating the factors or methods identified in (9) above into the complication prediction model and realizing a continuous feedback function. (11) Model accuracy improvement process A process to sequentially improve the accuracy of the complication prediction model by repeatedly performing steps (1) to (10) described above.

[0029] (Input data) The data used as input by the program described in this disclosure can be broadly categorized into "single-point-of-time data" and "time-series biometric information." This diverse information can be incorporated into the program through manual input, file upload, or integration with electronic medical records.

[0030] 1. Single-point-time data This primarily consists of static information determined preoperatively or during surgery, including, for example, the following. Additionally, time-series biological information, as described later, may be used as single-point-of-time data. Patient background information: Basic social background information of the patient. Patient physical information: Age, sex, height, weight, BMI, muscle mass, fat mass, nutritional status, ECOG-PS, ADL, lifestyle, physical activity level, muscle strength, physical function, motor function, cognitive function, frailty, sarcopenia, genetic information, etc. Comorbidities, medical history, surgical history, and medication information: Past or present conditions such as diabetes, hypertension, heart disease, arrhythmia, respiratory disease, neurological disorders, digestive disorders, kidney disease, urinary tract disorders, gynecological disorders, infections, thromboembolism, malignant tumors, dementia, collagen disease, allergies, smoking, and alcohol consumption. Also, a history of treatment for these conditions. Information on the disease being treated: Disease name, organ name, organ site, severity, preoperative treatment history (antibiotics, drainage, stent, mechanical ventilation, chemotherapy, radiation therapy, hormone therapy, etc.), tumor location, size, histopathological findings, stage, genetic information, etc. Surgical Information: Information about surgeries scheduled or performed prior to the operation. This includes whether emergency surgery was performed, the organs targeted, the surgical procedure, the extent of resection, the reconstruction method, the reconstruction site, the surgical approach (open surgery, craniotomy, retroperitoneal surgery, transanal surgery, laparoscopy, thoracoscopy, mediastinoscopy, arthroscopy, robotic surgery, etc.), the duration of the surgery, blood loss, degree of surgical cure, implants, surgical instruments, etc. Surgical images and videos may also be included.

[0031] 2. Time-series biological information This is dynamic information that tracks the temporal changes from before surgery to after surgery, and includes, for example, the following: Time-series biometric information may include, for example: 1) Vital data (pulse, heart rate, blood pressure, body temperature, respiratory rate, oxygen saturation [SpO2], etc.): When linked with wearable devices, continuous data from 1 minute to 24 hours is acquired. When manually entered or linked with electronic medical records, data is acquired as point-in-time data from 1 to 24 hours. 2) Other clinical data: Level of consciousness, urine output, drainage volume and characteristics, steps taken, activity level, sleep duration, food intake, etc., will be collected as time-series data at 1-24 hour intervals. 3) Blood biochemistry tests: Time-point data every 1 hour to 10 days (white blood cell count, neutrophil count, lymphocyte count, eosinophil count, hemoglobin level, hematocrit level, platelet count, CRP, procalcitonin, TP, Alb, T-Bil, D-Bil, AST, ALT, ALP, γ-GTP, ChE, LDH, AMY, BUN, Cr, Na, K, Cl, Ca, CK, blood glucose level, uric acid, iron, ferritin, PT, APTT, fibrinogen, D-dimer, etc.). 4) Examination information data: Radiographic images (X-ray, CT, MRI, etc.), nuclear medicine images (PET, SPECT, etc.), ultrasound examination information, endoscopic examination images and videos, electrocardiogram, heart rate variability, electroencephalogram, bioelectrical impedance information, etc. are acquired as time-series data at intervals of 1 to 24 hours. 5) Patient symptom information: Record the following symptoms every 1 to 24 hours: abdominal pain, abdominal distension, feeling of fullness, reflux symptoms, chest pain, sore throat, lower back pain, back pain, joint pain, muscle pain, wound pain, headache, pain, fever, convulsions, dizziness, lightheadedness, syncope, numbness, malaise, nausea, vomiting, loss of appetite, constipation, diarrhea, bloody stool, black stool, hematemesis, hemoptysis, hematuria, difficulty urinating, shortness of breath, frequent urination, oliguria, anuria, cough, shortness of breath, runny nose, palpitations, sweating, edema, facial flushing, visual impairment, hearing impairment, taste disorder, eczema, skin symptoms, hallucinations, auditory hallucinations, agitation, depression, irritability, irritability, insomnia, drowsiness, cognitive decline, memory impairment, etc.

[0032] These time-series data can be acquired continuously and automatically from measuring instruments and wearable devices, or intermittently by healthcare professionals. While time-series data from measuring instruments and wearable devices is useful, it is not essential. In this specification, "time series" refers to a concept that includes not only continuous and sustained data acquired from measuring instruments and wearable devices (e.g., heart rate acquired at intervals of 1-60 seconds, 1-60 minutes, 1-24 hours, and 1-7 days), but also intermittent and discrete data that can be analyzed as a temporal progression or pattern, such as data measured at regular intervals by healthcare professionals (e.g., every 1-60 minutes, every 1-24 hours, and every 1-7 days). The graph in Figure 1 shows an example of time-series changes in heart rate, blood pressure, and SpO2.

[0033] The single-point-of-time data, such as the patient's physical information mentioned above, can be incorporated into the program through manual input, file uploads, or integration with electronic medical records. This single-point-of-time data can be displayed, for example, in the first area of ​​Figure 2.

[0034] (Information to be output) The timing of the output of postoperative complication predictions is not limited to a single output before or after surgery; predictions can be calculated and output sequentially based on data input over time after surgery, although outputting only once before or after surgery is also acceptable. High accuracy is maintained because time-series information close to the output point is also incorporated. It is also possible to output complication risks and alerts in real time.

[0035] Furthermore, the program related to this disclosure can output information regarding complications, suggesting 1 to 50 recommended tests or treatments based on the calculated risk of developing or preventing the worsening of postoperative complications.

[0036] Furthermore, the program described herein may propose factors, change targets, and methods that can improve postoperative outcomes before surgery (for example, improvement of obesity, improvement of malnutrition, weight gain, improvement of diabetes, improvement of hypertension, abstinence from alcohol, smoking cessation, discontinuation / temporary suspension / change / initiation of medication, reduction of surgical scope, modification of surgical procedure, etc.).

[0037] (feedback) The program relating to this disclosure may have a feedback function that collects data, uses that data to identify new or highly accurate factors and methods for predicting complications, and improves the accuracy of complication prediction.

[0038] (flowchart) Figure 11 shows an illustrative flowchart of the program related to this disclosure. 1. Data entry (S1100): Patient information (age, target disease, etc.) and surgical information are entered into the program manually, via file upload, or through electronic medical record integration. Simultaneously, time-series biometric information such as vital data from wearable devices and measurements by healthcare professionals, as well as blood test results, are continuously or intermittently acquired and entered into the program. 2. Predictive execution (S1110): The user can initiate the prediction process at any time using buttons or other methods on the screen. 3. Analysis using models (S1120): The program aggregates the information accumulated up to that point (single-point-of-time data, time-series data). A trained AI model or other model analyzes not only the single-point-of-time data but also the trends and patterns in the time-series data, and calculates the probability of developing each of multiple postoperative complications (e.g., pneumonia, anastomotic leakage, etc.). 4. Risk calculation and warning determination (S1130): The system determines whether the calculated risk value exceeds a pre-set threshold (for example, "pneumonia risk of 30% or more"). If the threshold is exceeded, an alert is generated to draw the attention of healthcare professionals, patients, family members, or other people associated with the patient. 5. Display of results (S1140): The final complication risk is displayed in a user interface (UI) in an intuitive and easy-to-understand way, such as being listed in order of highest risk, or being color-coded and having changes in font and size according to the level of danger. 6. Suggestions for examinations and treatments (S1150): If necessary, 1 to 20 recommended tests and treatment options for high-risk complications will be presented.

[0039] (Input assistance) The program relating to this disclosure may further include functions that use generative artificial intelligence (LLM) when inputting or outputting data to provide input assistance for the input information, answer questions regarding the input or output information, or suggest optimal input parameters.

[0040] (Combination of predictive models) The program relating to this disclosure may include, for example, at least one of the following three, or possibly four, predictive models. Generally, the occurrence of each complication differs depending on the organ and surgical procedure. Therefore, the program relating to this disclosure may have a function (though not required) to categorize by combining organs and surgical procedures and list the high and low risks of complications. For example, 1 to 100 models are constructed to predict the risk of occurrence for 1 to 100 high-risk complications from a list of complications that differ depending on the organ and surgical procedure. The output may include: (1) the overall risk of complication occurrence (the risk of any complication occurring); (2) 1 to 100 patient-specific complications with a high risk of occurrence based on the model prediction results; (3) the top 1 to 100 specific complications with a high risk depending on the organ and surgical procedure; and (4) one, any, or all of the risk of death. Complications and deaths are targeted within 90 days after surgery, for example, within 30 days. Model 1: Calculate the risk of developing all postoperative complications. Model 2: Outputs complications (and their risks) that are predicted to occur frequently in the target patient. Model 3: This model calculates the risk of complications (e.g., intra-abdominal infection complications) that have been reported to have the highest risk of occurring in the target surgery in previous reports. Model 4: Calculate the risk of death after surgery.

[0041] The program disclosed herein may predict the risk of each complication from 1 to 30 models, including models of postoperative complications and postoperative mortality, representing the top 20 complications reported to have the highest risk of occurring in the target surgery, and output 1 to 10 of the highest risk-value complications to the user interface (UI).

[0042] For example, by combining multiple models as described above, both comprehensive and individual, specific risks can be presented to healthcare professionals or patients. As mentioned above, the types of complications are diverse.

[0043] Therefore, the postoperative complication prediction program relating to this disclosure is The first predictive model outputs the risk of developing all postoperative complications. A second predictive model that outputs complications predicted to occur frequently in the target patient, and A third predictive model that outputs the risk of developing complications that frequently occur in the aforementioned specific surgery, A fourth predictive model that outputs the risk of postoperative death. This may be characterized by outputting information regarding the aforementioned complications using one, two, three, or four predictive models selected from the group consisting of the following.

[0044] Furthermore, the postoperative complication prediction program relating to this disclosure may be characterized by using at least two or three models to output a predetermined number of high-risk complication risks.

[0045] Information on postoperative complications obtained as a result of model processing may include, but is not limited to, the presence or absence of complications, risk rate, date of onset, and severity. As mentioned above, postoperative complications are targeted within 90 days after surgery, for example, within 30 days, but most occur within 10 days.

[0046] The results of the model processing are output to, for example, the fourth area of ​​the UI. In the example in Figure 1, "High Risk (75%)" is displayed as sepsis, "Medium Risk (40%)" as pneumonia, and "Low Risk (15%)" as deep vein thrombosis. It is also possible to display them in order of risk (highest or lowest), and to distinguish them visually using color coding, font differences, size differences, etc.

[0047] The output of a complication prediction model may include, for example: 1. Overall risk of developing complications 2. Risk values ​​for complications ranked 1st to 10th in risk of onset, or those with an incidence rate of 10% to 50% or higher. 3. Risk values ​​for 1 to 10 types of high-risk complications specific to each surgical procedure. 4. Mortality risk value

[0048] (Alert) The program described herein may also include a function to issue an alert when the risk of complications exceeds a predetermined threshold, and a function to suggest multiple optimal tests and treatments according to the risk. Alerts may be notified, for example, by a display unit, sound, or vibration. Alternatively, a dedicated alert device may be used.

[0049] (Electronic medical record integration) The program related to this disclosure may also include a function to output input information and model processing results to an electronic medical record.

[0050] (User Interface) Figure 1 shows an example of the user interface (UI) of this embodiment. As illustrated in Figure 1, the UI may mainly consist of an area for inputting and displaying patient information, an area for displaying time-series data of biometric information acquired from wearable devices, etc., in a graph, and an area for outputting the complication risk calculated by model processing, but is not limited to this configuration.

[0051] Other example screen designs are shown in Figures 2-10. The first area displays single-point-of-time data (patient physical information, comorbidity information, target disease information, surgical information, etc.). The second area displays time-series biometric information (vital data, other clinical data, blood test results, patient symptom information, time-series graphs, etc.). The third area contains the output button for the model results. The fourth area displays risk values ​​for 1 to 20 types of complications. Visual effects such as ranking, color coding, and emphasis using fonts and sizes may be added. Furthermore, alerts may be issued for high-risk complications.

[0052] The user interface (UI) may input at least one of the following into the first domain: patient physical information, patient symptom information, comorbidity information, preoperative treatment, target disease information, blood test information, radiographic image information, ultrasound information, surgical information, etc.; input at least one time-series data of vital signs such as pulse, heart rate, blood pressure, body temperature, respiratory rate, and SpO2 (blood oxygen saturation) into the second domain; and output the risk of postoperative complications into the fourth domain. It is also possible to incorporate radiographic images (e.g., preoperative CT images), ultrasound information, electrocardiograms, heart rate variability, electroencephalograms, bioelectrical impedance information, and surgical images or videos as reference information, but in that case, time-series display in the UI is not mandatory. A button to output the model results may be placed in the third domain. The UI may also model-predict three (or two or more) types of complications and display them in a manner corresponding to the level of risk (or whether certain requirements are met). Here, data input is possible by either manual input or import of external data, and is not limited to a specific method. Vital data can be entered manually or migrated from wearable devices. Furthermore, the UI may include a function to display warnings when the risk of complications is high. Display methods include, but are not limited to, outputting data in order of risk, color-coding, or other visual distinctions.

[0053] Furthermore, the program disclosed herein can operate on smartphone apps, tablets, and desktop computers, and may include a function to output input information externally. The program disclosed herein may also enable the input and output of data and models through interconnection with electronic medical records.

[0054] (method) One aspect of this disclosure relates to a method for predicting postoperative complications. The method according to this disclosure includes, for example, the steps of: acquiring time-series biological information from a patient who has undergone surgery from before surgery to a predetermined period after surgery; and inputting the acquired time-series biological information into a trained artificial intelligence model to acquire information about postoperative complications output by the model. The method according to this disclosure can be performed on a computer.

[0055] More specifically, the method relating to this disclosure involves, for example, sequentially performing the following steps. 1. Patient Information Input Steps Basic information about the patient to be predicted (age, sex, comorbidities, etc.) and information about the surgery performed (surgical procedure, surgery time, blood loss, etc.) are entered into the program. This input is performed either through automatic integration from the electronic medical record system or manually by healthcare professionals.

[0056] 2. Steps for acquiring time-series biological information The system continuously or intermittently acquires and records vital signs such as heart rate, blood pressure, respiratory rate, and SpO2, as well as blood test results and patient symptoms, from wearable devices, bedside monitors, and electronic medical records. This step also includes the manual input of periodic measurements by healthcare professionals, patients, relatives, and patient contacts.

[0057] 3. Prediction execution step Medical professionals such as doctors and nurses, as well as patients, relatives, and those related to patients, can instruct the program to perform predictive actions on a computer screen, but the execution is not limited to these groups. In response to the instructions, the program collects the entered patient information and the time-series biometric data accumulated up to that point in a single operation.

[0058] 4. Analysis and Risk Calculation Steps The program inputs some or all of the collected information into a trained artificial intelligence model. The AI ​​model analyzes the patterns and trends in the data and calculates the future risk of developing specific postoperative complications, such as "intraperitoneal infection" and "pneumonia," as a probability (e.g., 75%).

[0059] 5. Results presentation and warning step The calculated risk of each complication is displayed on the screen in a list format, ordered from highest to lowest risk. The risk levels are visually indicated, such as "high risk (red)" or "medium risk (yellow)." If any risk exceeds a pre-set threshold, an alert notification (warning display or sound) is automatically triggered to draw the attention of healthcare professionals.

[0060] Through these steps, healthcare professionals, patients, relatives, and those involved with patients can use objective data to detect early signs of complications and consider preventive testing and treatment interventions.

[0061] (Method for generating a postoperative complication prediction model) One aspect of this disclosure relates to a method for generating a postoperative complication prediction model. The model generation method according to this disclosure includes, for example, the steps of: obtaining time-series biological information from before surgical procedures to a predetermined period postoperatively, and actual data regarding the occurrence of complications within 90 days postoperatively for multiple past patient cases; and training a machine learning model that predicts postoperative complications, using the time-series biological information as input and the actual data as ground truth labels.

[0062] (Machine learning model) One aspect of this disclosure relates to a machine learning model. The machine learning model relating to this disclosure may be, for example, a trained machine learning model generated by the method described above. Therefore, the machine learning model relating to this disclosure may be, for example, a trained machine learning model generated by a method that includes the steps of: acquiring time-series biological information for multiple past patient cases from before surgical procedures to a predetermined period after surgery, and actual data regarding the occurrence of complications within 90 days after surgery; and training a machine learning model that predicts postoperative complications using the time-series biological information as input and the actual data as ground truth labels.

[0063] (Device) One aspect of this disclosure relates to a postoperative complication prediction device. The device according to this disclosure includes, for example, an acquisition unit that acquires time-series biological information of a patient who has undergone surgery from before surgery to a predetermined period after surgery, and a processing unit that inputs the acquired time-series biological information into a trained artificial intelligence model and acquires information on postoperative complications output by the model.

[0064] The device relating to this disclosure may specifically be configured as a computer system equipped with a CPU (Central Processing Unit), RAM (Main Memory), auxiliary storage devices such as SSDs and hard disk drives, display devices such as liquid crystal displays, input devices such as keyboards and mice, and a network interface. The device relating to this disclosure may also be a wearable device.

[0065] In this case, the auxiliary storage device stores the operating system (OS) along with the postoperative complication prediction program according to this disclosure. The CPU then executes this program, causing the entire computer to function as a postoperative complication prediction device, fulfilling the roles of the acquisition unit and processing unit mentioned above.

[0066] More specifically, this device may include, for example, the following components: 1. Control Unit: Consists of a CPU and other components, and provides overall control over the operation of the entire device. 2. Memory Unit: Composed of RAM, ROM, SSD, etc. It stores the postoperative complication prediction program, trained artificial intelligence models, input patient data, and output prediction results. 3. Acquisition Unit: This unit consists of a network interface, keyboard, touch panel, etc. It receives time-series biometric information from electronic medical record systems and wearable devices via the network, and also accepts manual input from medical professionals. 4. Processing Unit: This is a functional unit realized by the control unit executing a program stored in the memory unit. It inputs time-series biological information and other patient information acquired by the acquisition unit into a trained artificial intelligence model stored in the memory unit, calculates the risk of developing complications, and outputs it. 5. Output Unit: This unit consists of an LCD display, speakers, network interface, etc. It displays the complication risk calculated by the processing unit on the screen, emits a warning sound if the risk exceeds a predetermined threshold, and transmits the results to the electronic medical record system. These components work together in coordination to enable the prediction of postoperative complications.

[0067] (Functional Configuration) Figure 12 is a block diagram showing the functional configuration of a postoperative complication prediction device 1200 according to one embodiment of the present disclosure. The device 1200 mainly comprises a control unit 1210 that comprehensively controls the operation of the entire device, a storage unit 1220 that stores various data and programs, an acquisition unit 1230 that acquires information from the outside, a processing unit 1240 that performs information processing, and an output unit 1250 that outputs processing results to the outside.

[0068] The control unit (1210) consists of a CPU (Central Processing Unit) and other components, and comprehensively controls the operation of each component of the entire device by reading and executing a postoperative complication prediction program from the memory unit 1220.

[0069] The memory unit (1220) consists of RAM, ROM, SSD, hard disk drive, etc. It stores the OS (operating system), the postoperative complication prediction program related to this disclosure, a trained artificial intelligence model, patient data acquired by the acquisition unit 1230, and prediction results calculated by the processing unit 1240.

[0070] Acquisition unit (1230): Composed of a network interface, keyboard, mouse, touch panel, etc. It receives time-series biometric information and basic patient information from external sources such as electronic medical record systems and wearable devices via the network. It also accepts manual input from medical professionals using a keyboard, etc.

[0071] Processing Unit (1240): This is a functional unit realized by the control unit 1210 executing a program. The time-series biological information and other patient information acquired by the acquisition unit 1230 is input into a trained artificial intelligence model stored in the memory unit 1220, and the risk of developing postoperative complications is calculated and output.

[0072] Output unit (1250): Consists of an LCD display, speaker, network interface, etc. It displays the complication risk calculated by the processing unit 1240 on the display and emits a warning sound if the risk exceeds a predetermined threshold. It also has the function of sending the prediction results to the electronic medical record system.

[0073] These components work in coordination with each other under the control of the control unit 1210, thereby enabling early prediction of postoperative complications.

[0074] (Information flow) Figure 13 is a schematic diagram illustrating the flow of information centered on a postoperative complication prediction program 1300 according to one embodiment of this disclosure. The flow of information in this program can be broadly composed of three stages: an input stage, a processing stage, and an output stage.

[0075] Input stage (1310): This program aggregates information from diverse data sources. Specifically, wearable devices worn by patients continuously input time-series vital signs such as heart rate and SpO2. Structured data such as basic patient information, blood test data, and surgical records are also linked from the hospital's electronic health record (EHR) system. Furthermore, unstructured data such as symptoms and observations from the patient themselves or healthcare professionals may be manually entered.

[0076] Processing stage (1320): The diverse input data is first collected in the program's internal data aggregation and storage unit, where it is organized and integrated chronologically. Then, the core function, the AI ​​prediction model, analyzes this integrated data. The AI ​​prediction model evaluates not only the individual values ​​of the data but also their temporal progression and the correlation between multiple pieces of information, calculating the risk of a specific postoperative complication occurring as a probability.

[0077] Output stage (1330): The analysis results from the AI ​​predictive model are formatted and output on the user interface (UI). On PC, smartphone, and tablet screens, risk values ​​for each complication, time-series graphs of vital signs, and warnings (alerts) are displayed if the risk exceeds a threshold. Based on this information, users (healthcare professionals) can assess the patient's condition and make clinical decisions regarding subsequent tests and treatments. Patients, relatives, and those involved with the patient can also use this information to suggest further tests and treatments to healthcare professionals. Furthermore, the calculated risk information and alert records may be written back to the electronic health record (EHR) system for future reference and auditing, and may be stored as part of the official medical record.

[0078] (system) One aspect of this disclosure relates to a postoperative complication prediction system. The system according to this disclosure may include, for example, at least one measuring device that acquires time-series biological information of a patient who has undergone surgery from before surgery to a predetermined period after surgery, and a processing device configured to receive the time-series biological information from the measuring device and to perform a model processing step that outputs information about postoperative complications that will occur within 90 days after surgery based on the time-series biological information.

[0079] (Recording medium) One aspect of this disclosure relates to a computer-readable recording medium on which a postoperative complication prediction program is recorded. Such a recording medium is not particularly limited as long as it stores information by electrical, magnetic, optical, mechanical, or chemical means and is computer-readable, and includes, for example, hard disk drives (HDDs), solid-state drives (SSDs), CD-ROMs, DVDs, Blu-ray® Discs, USB memory sticks, SD cards, and the like.

[0080] (Interpretation of terms) Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those generally understood by those skilled in the art to which the invention pertains. Any methods and materials similar or equivalent to those described herein may be used for carrying out or testing the invention, but several possible and preferred methods and materials are described herein. All publications referenced herein are incorporated herein by reference, and the methods and / or materials cited in relation to these publications are disclosed and described herein. In the event of any conflict, this disclosure shall prevail over the disclosure of the incorporated publication.

[0081] Where a range of values ​​is given, unless the context clearly indicates otherwise, each intermediate value between the upper and lower limits of that range, up to one-tenth of the lower limit unit, is also specifically disclosed. Each smaller range between any given or intermediate value within the given range and any other given or intermediate value within that given range is also included in this disclosure. The upper and lower limits of these smaller ranges may be independently included in or excluded from that range, and each range that includes either, either, or both of the limit values ​​in the smaller range is also included in the invention, but the limit values ​​specifically excluded in the given range are reserved. Where a given range includes one or both of the limit values, a range that excludes either or both of the included limit values ​​is also included in the invention. The term “about” with respect to numerical values ​​means within 5%.

[0082] The embodiments described herein are intended to be illustrative only, and those skilled in the art will be able to make numerous modifications and alterations without departing from the spirit of the invention. Certain modifications and alterations may yield satisfactory results, though not optimal. All such modifications and alterations are intended to fall within the scope of the invention as defined by the appended claims. Furthermore, any combination of the components disclosed herein, and any conversion of the expressions of this disclosure between methods, apparatus, computer programs, systems, data structures, trained machine learning models, recording media, etc., are also valid embodiments of this disclosure. Therefore, the details described with respect to each of the methods, apparatus, computer programs, systems, data structures, trained machine learning models, recording media, etc., may be applied as appropriate to each of the methods, apparatus, computer programs, systems, data structures, trained machine learning models, recording media, etc. [Examples]

[0083] (Example 1: Model Construction) A complication prediction model was constructed by training it with data from 1100 surgeries involving malignant tumors (cancer, GIST, sarcoma, etc.) and benign diseases (gastrointestinal perforation, intestinal obstruction, cholecystitis, etc.) in the gastrointestinal tract and hepatobiliary-pancreatic region. A separate validation set of 50 surgeries was used.

[0084] Using single-time period data such as age, sex, comorbidities, and surgical procedure, we constructed predictive models for all postoperative complications and intra-abdominal infectious complications using logistic regression analysis. In the validation set, the sensitivity, AUC, and F1 score were 0.520, 0.655, and 0.429 for all complications, and 0.580, 0.638, and 0.276 for intra-abdominal infectious complications, indicating that high-precision prediction was not achieved.

[0085] Next, a logistic regression analysis model was constructed using time-series data, including pulse rate, heart rate, body temperature, respiratory rate, SpO2, urine output, and blood test data (white blood cell count, CRP, creatinine level, blood glucose level, albumin level, etc.) from before surgery to 3 days post-surgery. This model was then combined with the aforementioned single-time-point data model to create a mixed model. As a result, the prediction of all complications showed a sensitivity of 0.833, an AUC of 0.729, and an F1 score of 0.652, while the prediction of intraperitoneal infectious complications showed improved accuracy with a sensitivity of 0.857, an AUC of 0.801, and an F1 score of 0.632. Furthermore, when a random forest model was constructed using similar time-series data and mixed with the aforementioned single-time-point data model, the prediction of all complications showed a sensitivity of 0.670, an AUC of 0.731, and an F1 score of 0.649, while the prediction of intraperitoneal infectious complications showed a sensitivity of 0.857, an AUC of 0.834, and an F1 score of 0.571.

[0086] In addition, using CT scan image data from the chest to the abdomen and pelvis (approximately 130 images per patient) performed before surgery, a 2D-CNN model based on ResNet18 was constructed as a Deep Learning model using 5-fold cross-validation. A three-part hybrid model was then created, incorporating single-time point-in-time data and time-series data. The prediction of intra-abdominal infectious complications showed a sensitivity of 0.857, an AUC of 0.854, and an F1 score of 0.571, demonstrating the highest accuracy among all models.

[0087] (Example 2: Anticipated Use Case 1) Among the complications of gastric cancer surgery, the highest risk complications, in descending order, are intra-abdominal abscess, pneumonia, anastomotic leakage, pancreatic fistula, wound infection, intestinal obstruction, enteritis, bleeding, urinary tract infection, biliary tract infection, catheter-related infection, lymph leakage, liver damage, kidney damage, thromboembolism, allergy, and drug-induced organ damage.

[0088] An example of the output from the model processing for gastric cancer surgery is as follows: 1. Risk of developing all complications: 55% 2. Complications with a high risk of onset as predicted by the model: 1st: Cerebral infarction 45% 2nd: Enteritis 25% 3. Complications with high risk values ​​specific to the surgical procedure: Intra-abdominal abscess 2%, Pneumonia 1%, Anastomotic leakage 0.4% 4. Risk of surgery-related death: 0.03%

[0089] (Example 3: Anticipated Use Case 2) High-risk complications associated with surgery for pneumothorax of the lung include pneumonia, atelectasis, pneumothorax, arrhythmia, empyema, wound infection, tracheal stump fistula, bleeding, respiratory failure, acute respiratory distress syndrome, urinary tract infection, catheter-related infection, myocardial infarction, angina pectoris, heart failure, cerebral infarction, cerebral hemorrhage, thromboembolism, arterial dissection, edema, allergies, and drug-induced organ damage.

[0090] An example of the output from the model processing for pneumothorax surgery is as follows: 1. Risk of developing all complications: 35% 2. Complications with a high risk of onset as predicted by the model: 1st place: Arrhythmia 20% 3. Complications with high risk values ​​specific to the surgical procedure: Pneumonia 3%, Atelectasis 1% 4. Risk of surgery-related death: 0.1% [Explanation of Symbols]

[0091] 1200... Postoperative complication prediction device 1210... Control Unit 1220...Storage section 1230...Acquisition part 1240... Processing Unit 1250... Output section 1300...Postoperative Complication Prediction Program 1310... Input stage 1320... Processing stage 1330... Output stage

Claims

1. A step of obtaining time-series biological information of a patient who has undergone surgery from before surgery to a predetermined period after surgery, as well as surgical information including at least one of the surgical organ, surgical procedure, extent of resection, reconstruction method, and surgical approach of the patient. A model processing step that outputs information on postoperative complications, including at least one selected from the group consisting of all postoperative complications, intra-abdominal infection complications, abscess, ascites, pleural effusion, pneumothorax, anastomotic leakage, pancreatic fistula, bleeding, respiratory failure, urinary tract infection, biliary tract infection, catheter-related infection, vasculitis, intestinal obstruction, enteritis, cerebral infarction, cerebral hemorrhage, seizures, liver dysfunction, and thromboembolism, based on the acquired time-series biological information and surgical information. Have the computer run it, A postoperative complication prediction program characterized by its ability to predict complications that will occur within 90 days after surgery.

2. The postoperative complication prediction program according to claim 1, characterized in that the model processing step further uses, in addition to the time-series biological information, at least one piece of information selected from the group consisting of patient physical information, comorbidity information, treatment target disease information, and surgical information, which are data from a single point in time, as input, and outputs information regarding the complication.

3. The postoperative complication prediction program according to claim 1, characterized in that the model processing step further uses surgical images or videos as input, extracts spatial features from the images or videos using machine learning techniques including deep learning, and outputs information about the complications by combining the time-series features extracted from the time-series biological information with the spatial features.

4. A first predictive model that outputs the risk of developing all postoperative complications based on the surgical information, A second predictive model outputs complications that are predicted to occur frequently in the target patient based on the aforementioned surgical information. A third predictive model that outputs the risk of developing complications that frequently occur in specific surgeries based on the aforementioned surgical information. A fourth predictive model that outputs the risk of death based on the aforementioned surgical information, A postoperative complication prediction program according to claim 1, characterized in that it outputs information about the complication using one, two, three, or four predictive models selected from the group consisting of the following.

5. The postoperative complication prediction program according to claim 1, characterized in that the model processing step outputs a predetermined number of high-risk complication risks using at least two models.

6. The postoperative complication prediction program according to claim 1, characterized in that, as information regarding the aforementioned complication, it issues a warning when the risk of developing the complication exceeds a predetermined threshold.

7. The postoperative complication prediction program according to claim 1, characterized in that, as information regarding the aforementioned complications, it outputs information suggesting 1 to 50 recommended tests or treatments according to the risk of onset calculated for the purpose of reducing the risk of developing or preventing the worsening of postoperative complications.

8. The postoperative complication prediction program according to claim 1, further comprising a function that uses generative artificial intelligence to provide input support for the acquired information, answer questions regarding the input or output information, or suggest optimal input parameters.

9. A postoperative complication prediction program according to claim 1, characterized in that predicted values ​​are calculated over time, reflecting the input of data over time after surgery.

10. A postoperative complication prediction program according to claim 1, characterized in that it outputs complication risk and / or alerts in real time.

11. A postoperative complication prediction program according to claim 1, characterized by proposing factors, change targets, or methods that can improve postoperative outcomes before surgery.

12. A postoperative complication prediction program according to claim 1, characterized in that it collects data, uses that data to identify new or highly accurate factors or methods useful for predicting complications, and has a feedback function to improve the accuracy of complication prediction.

13. The postoperative complication prediction program according to claim 1, wherein a display device connected to a computer displays, as a user interface, a first area for inputting single-point-of-time data, a second area for inputting time-series data, a third area for accepting operations to perform analysis of a postoperative complication prediction model, and a fourth area for outputting the results of the analysis of the postoperative complication prediction model.

14. The postoperative complication prediction program according to claim 13, wherein the second domain includes a manual data input function, a data linkage function with measuring instruments, and a function for displaying the temporal changes of the input time-series data as a chart, and the fourth domain has an output format corresponding to the risk level of postoperative complications and includes a warning function that issues a warning when the risk level exceeds a predetermined threshold.

15. A step of obtaining time-series biological information of a patient who has undergone surgery from before surgery to a predetermined period after surgery, as well as surgical information including at least one of the surgical organ, surgical procedure, extent of resection, reconstruction method, and surgical approach of the patient. The steps include inputting the acquired time-series biological information and surgical information into a trained artificial intelligence model to obtain information on postoperative complications, including at least one selected from the group consisting of all postoperative complications, intra-abdominal infection complications, abscess, ascites, pleural effusion, pneumothorax, anastomotic leakage, pancreatic fistula, bleeding, respiratory failure, urinary tract infection, biliary tract infection, catheter-related infection, vasculitis, intestinal obstruction, enteritis, cerebral infarction, cerebral hemorrhage, seizures, liver dysfunction, and thromboembolism, and obtaining information on postoperative complications output by the model. A method for predicting postoperative complications that occur within 90 days after surgery, including those mentioned above.

16. An acquisition unit that acquires time-series biological information of a patient who has undergone surgery from before surgery to a predetermined period after surgery, as well as surgical information including at least one of the surgical organ, surgical procedure, resection area, reconstruction method, and surgical approach of the patient. A processing unit inputs the acquired time-series biological information and surgical information into a trained artificial intelligence model and acquires information on postoperative complications, including at least one selected from the group consisting of all postoperative complications, intra-abdominal infection complications, abscess, ascites, pleural effusion, pneumothorax, anastomotic leakage, pancreatic fistula, bleeding, respiratory failure, urinary tract infection, biliary tract infection, catheter-related infection, vasculitis, intestinal obstruction, enteritis, cerebral infarction, cerebral hemorrhage, seizures, liver dysfunction, and thromboembolism. A device for predicting postoperative complications that occur within 90 days after surgery, including those mentioned above.

17. A computer-readable recording medium that records the postoperative complication prediction program described in claim 1.

18. A measuring device that acquires time-series biological information from before surgery to a predetermined period after surgery in patients who have undergone surgical procedures, A processing device configured to receive the time-series biological information from the measurement device and to perform a model processing step that outputs information on postoperative complications that occur within 90 days postoperatively, including at least one selected from the group consisting of all postoperative complications, intra-abdominal infection complications, abscess, ascites, pleural effusion, pneumothorax, anastomotic leakage, pancreatic fistula, bleeding, respiratory failure, urinary tract infection, biliary tract infection, catheter-related infection, vasculitis, intestinal obstruction, enteritis, cerebral infarction, cerebral hemorrhage, seizures, liver dysfunction, and thromboembolism, based on the time-series biological information and surgical information including at least one of the patient's surgical target organ, surgical procedure, resection range, reconstruction method, and surgical approach. A postoperative complication prediction system characterized by comprising the following features.

19. The steps include obtaining, for multiple past patient cases, chronological biological information from before surgery to a predetermined period post-surgery, surgical information including at least one of the patient's surgical target organ, surgical procedure, resection extent, reconstruction method, and surgical approach, and performance data regarding the occurrence of complications within 90 days post-surgery, The steps involve training a machine learning model to predict postoperative complications, which include at least one selected from the group consisting of all postoperative complications, intra-abdominal infection complications, abscess, ascites, pleural effusion, pneumothorax, anastomotic leakage, pancreatic fistula, bleeding, respiratory failure, urinary tract infection, biliary tract infection, catheter-related infection, vasculitis, intestinal obstruction, enteritis, cerebral infarction, cerebral hemorrhage, seizures, liver dysfunction, and thromboembolism, using the aforementioned time-series biological and surgical information as input and the aforementioned actual data as the ground truth label. A method for generating a postoperative complication prediction model, characterized by including the following:

20. The postoperative complication prediction program according to claim 1, wherein the surgical procedure is any of the following: gastrointestinal surgery, urological surgery, respiratory surgery, head and neck surgery, breast surgery, cardiac surgery, vascular surgery, neurosurgery, or emergency surgery.

Images