Inpatient frailty identification system

Abstract

The inpatient frailty identification system disclosed in this invention includes a database and a processing module. The database stores a first medical record of a patient to be analyzed and a frailty classification model. The first medical record contains multiple first medical records of the patient at different time points during hospitalization. The processing module, using an analysis time point as a reference, reads the first medical records of the patient to be analyzed that are recorded earlier than and closest to the analysis time point. It extracts a first key feature from the selected first medical records and inputs this first key feature into the frailty classification model to analyze and obtain a frailty identification result for the patient to be analyzed. Accordingly, the inpatient frailty identification system can quickly identify whether a hospitalized patient is in a frail state, assisting medical staff in taking necessary care measures as early as possible.

Description

[Technical Field]

[0001] This invention relates to detection technology, and in particular to a system for identifying the weakness of hospitalized patients. [Previous Technology]

[0002] Frailty is a significant issue affecting the health and quality of life of older adults. It typically refers to an abnormal aging process and is a reversible pathological state associated with various risk factors and adverse outcomes. Frailty is characterized by reduced health reserves, decreased physiological function, and impaired self-regulation, thus increasing individual vulnerability. The increased risk of adverse outcomes due to frailty, such as physical disability, falls, hospitalization, institutionalization, and premature death, necessitates public attention and concern.

[0003] Therefore, the diagnosis of frailty is particularly important, generally referring to the process of assessing the physiological decline and health risks of older adults using various tools or indicators. However, there is currently no universally accepted standard to define and measure frailty, but some commonly used methods exist, such as Fried's frailty phenotype, the frailty index, and the Clinical Frailty Scale. These methods each have their advantages and disadvantages and are suitable for different occasions and purposes. Therefore, the diagnosis of frailty should include assessments of physical, cognitive, emotional, social, and mental aspects, taking into account individual characteristics and needs, which makes the assessment process quite lengthy, and there is still room for improvement in the accuracy of the assessment results. [Summary of the Invention]

[0004] Therefore, the main objective of this invention is to provide a frailty identification system for hospitalized patients, which can quickly identify whether a hospitalized patient is in a frail state and assist medical staff in taking necessary care measures as early as possible.

[0005] Therefore, in order to achieve the above objectives, the inpatient frailty identification system disclosed in this invention includes a database and a processing module. The database stores a first medical record of a patient to be analyzed and a frailty classification model. The first medical record contains multiple first medical data recorded at different time points during the hospitalization of the patient to be analyzed. The processing module reads the first medical data of the patient to be analyzed from the first medical record, which is recorded earlier than the analysis time point and is closest to the analysis time point, based on an analysis time point. It extracts a first key feature from the selected first medical data and then inputs the first key feature into the frailty classification model to analyze a frailty identification result for the patient to be analyzed.

[0006] Furthermore, the analysis time point is N days after the patient to be analyzed is admitted to the hospital, where N is a positive integer; the first key feature includes gender, age, latest blood glucose (Glu_latest), latest blood potassium (K_latest), latest blood sodium (NA_latest), latest blood creatinine (CREAT_latest), latest white blood cell count (WBC_latest), latest neutrophil count (NEUT_latest), latest C-reactive protein (CRP_latest), latest blood urea nitrogen (BUN_latest), latest heme (Hgb_latest), latest glomerular filtration rate (EGFR_latest), latest platelet count (PLT_latest), nasogastric tube (NG), urinary catheter (Foley), restraint (Restraint), fasting (NPO), number of hospitalizations (H_times), number of emergency room visits (ER_times), frailty index (FI), and emergency triage level (Triage).

[0007] In one embodiment, the database further stores a plurality of second medical records of different hospitalized patients. Each second medical record contains a plurality of second medical data collected at different time points during the hospitalization of each hospitalized patient, and the health status of each hospitalized patient, which is either non-frail or frail. The processing module reads the second medical data recorded last in chronological order from each second medical record, and extracts a second key feature from the selected second medical data. The second key features and the health status of the hospitalized patients are then used as training data for constructing the frailty classification model. The second key features include gender, age, and latest blood glucose level (Glu_latest). Latest serum potassium (K_latest), latest serum sodium (NA_latest), latest serum creatinine (CREAT_latest), latest white blood cell count (WBC_latest), latest neutrophil count (NEUT_latest), latest C-reactive protein (CRP_latest), latest blood urea nitrogen (BUN_latest), latest heme (Hgb_latest), latest glomerular filtration rate (EGFR_latest), latest platelet count (PLT_latest), nasogastric tube (NG), urinary catheter (Foley), restraint (Restraint), fasting (NPO), number of hospitalizations (H_times), number of emergency room visits (ER_times), frailty index (FI), and emergency triage level (Triage).

[0008] In one embodiment, each of the second key features is further categorized into a laboratory dataset, a treatment order dataset, a basic dataset, and other datasets, wherein the laboratory dataset includes the latest blood glucose level (Glu_latest). The dataset includes the latest serum potassium (K_latest), serum sodium (NA_latest), serum creatinine (CREAT_latest), latest white blood cell count (WBC_latest), latest neutrophil count (NEUT_latest), latest C-reactive protein (CRP_latest), latest blood urea nitrogen (BUN_latest), latest heme (Hgb_latest), latest glomerular filtration rate (EGFR_latest), and latest platelet count (PLT_latest). The treatment order dataset includes nasogastric tube (NG), urinary catheter (Foley), restraint (Restraint), and fasting (NPO). The basic dataset includes gender (Gender) and age (Age). Other datasets include hospitalizations (H_times), emergency room visits (ER_times), frailty index (FI), and emergency triage level (Triage).

[0009] In one embodiment, in order to make the trained frailty classification model more accurate, the processing module also uses a data cleaning program to determine whether the second key feature is abnormal, and deletes the second key feature and health status of the hospitalized patient to which the second key feature is determined to be abnormal from the training data of the frailty classification model.

[0010] In one embodiment, the data cleaning procedure refers to the processing module checking whether the latest blood glucose (Glu_latest), latest blood potassium (K_latest), latest blood sodium (NA_latest), latest blood creatinine (CREAT_latest), latest white blood cell (WBC_latest), latest neutrophil (NEUT_latest), latest C-reactive protein (CRP_latest), latest blood urea nitrogen (BUN_latest), latest heme (Hgb_latest), latest glomerular filtration rate (EGFR_latest), and latest platelet (PLT_latest) in the laboratory dataset of each of the second key features are missing more than a predetermined number, and defining the second key feature missing more than the predetermined number as abnormal.

[0011] In one embodiment, the predetermined quantity is 6.

[0012] In one embodiment, the data cleaning procedure refers to the processing module confirming whether the number of days of hospitalization for each hospitalized patient is less than the analysis time point, and defining the second key feature extracted from the second medical record of the hospitalized patient whose number of days of hospitalization is less than the analysis time point as abnormal.

[0013] In one embodiment, the data cleaning procedure causes the processing module to check whether the date on which each hospitalized patient completed the health status assessment is greater than the analysis time point, and defines the second key feature extracted from the second medical record of the hospitalized patient whose assessment date is greater than the analysis time point as abnormal.

[0014] In one embodiment, N is 3.

[0015] In one embodiment, the weakened classification model is selected from a group consisting of Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), k-nearest neighbor (KNN), Random Forest (RF) algorithm and Logistic Regression (LR).

Implementation Method

[0017] The following will explain the terms or expressions used in this invention. If not listed below, they will be interpreted based on references recognized by those skilled in the art, such as dictionaries, encyclopedias, literature, or common knowledge. For example, the terms "average," "variance," "maximum," "minimum," and "latest" used in this invention describe the statistical characteristics of data and the attributes of different values. These terms are common in the fields of statistics and data analysis and are well-known to those skilled in the art.

[0018] Please refer to Figure 1. The inpatient frailty identification system disclosed in the preferred embodiment of the present invention includes a database 10 and a processing module 20.

[0019] The database 10 is a structured collection of information or data, which is stored in a predetermined manner, such as electronically in a computer system, hard drive, or cloud space. Furthermore, the database 10 is a non-volatile or volatile storage device pre-installed in a computing device, such as a hard drive (HDD), solid-state drive (SSD), phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), flash memory disk, read-only memory (ROM), or random access memory (RAM), used to store a first medical record, multiple second medical records, and a weak classification model. The source of this stored data may be direct input or obtained by concatenating with other external databases. The computing device can be, but is not limited to, a personal computer (PC), laptop, tablet, smartphone, or other device with computing or computational capabilities. The minimum requirement for the central processing unit (CPU) of this device is typically a multi-core, high-speed CPU, such as an Intel Core i5, AMD Ryzen 5, or higher. However, as the complexity of artificial intelligence tasks increases and the scale of data expands, more advanced CPUs or processors specifically designed to accelerate deep learning tasks (such as NVIDIA GPUs) may be more suitable.

[0020] The first medical record refers to the relevant records of a patient to be analyzed. In this example, the patient to be analyzed is an elderly person, i.e., 65 years old. However, in other possible cases, the age of the patient to be analyzed is not limited to this. Furthermore, the first medical record includes multiple first medical records, each of which is relevant data or information recorded at different time points during the hospitalization of the patient to be analyzed. Moreover, the time range of the recorded or data values ​​is between six months before hospitalization and three days after hospitalization.

[0021] Each of these second medical records corresponds to the relevant records of a hospitalized patient. In this example, each of these hospitalized patients is an elderly person admitted to a geriatric ward, and a total of 7421 second medical records were collected as reference materials for subsequent work. Furthermore, each of these second medical records contains multiple sets of second medical information and health status. The second medical information consists of relevant data or information collected at different time points during the hospitalization of each hospitalized patient. Moreover, the health status is the result of a comprehensive geriatric assessment of each hospitalized patient's diagnosis based on Fried's frailty phenotype, i.e., the health status is either non-frail or frail.

[0022] This weakened classification model is constructed using computational logic or algorithms known in the technical field to which this invention pertains, such as Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), k-nearest neighbor (KNN), Random Forest (RF), and Logistic Regression (LR). After testing different algorithms, Random Forest and Support Vector Machine achieved better prediction results, as shown in Tables 1 to 9. Furthermore, SHAP (SHapley Additive exPlanations) graphs and partial dependency graphs can be used to explore the interpretability of each feature in the prediction model.

[0023] The processing module 20 is a program or software that is pre-installed in the database 10 and used by the computing device to run, and can perform operations such as "calculation", "analysis", "judgment", "comparison", "matching" and "operation" on the received data. In this example, the processing module 20 executes a model building program, a data cleaning program S2, a model performance evaluation program S3 and a weakness identification program, which are described in detail below.

[0024] First, in the model construction program, as shown in Figure 25, the processing module 20 reads the second medical data recorded in the second medical records according to the last time in the time sequence, and extracts a second key feature from the selected second medical data, performs the data cleaning program S2, and then uses the cleaned second key features and the health status of the hospitalized patients as training data for constructing the frailty classification model, and stores the trained frailty classification model in the database 10. The second key feature includes gender, age, latest blood glucose (Glu_latest), latest blood potassium (K_latest), latest blood sodium (NA_latest), latest blood creatinine (CREAT_latest), latest white blood cell count (WBC_latest), latest neutrophil count (NEUT_latest), latest C-reactive protein (CRP_latest), latest blood urea nitrogen (BUN_latest), latest heme (Hgb_latest), latest glomerular filtration rate (EGFR_latest), latest platelet count (PLT_latest), nasogastric tube (NG), urinary catheter (Foley), restraint (Restraint), fasting (NPO), number of hospitalizations (H_times), number of emergency room visits (ER_times), frailty index (FI), and emergency triage level (Triage). These features included in the second key feature were selected. In particular, Figure 2 illustrates the impact of these 21 features on the detection results of hospitalized frailty. Furthermore, as shown in Figures 4 to 24, the dependency plots (dependence_plots) are generated using the 21 features contained in this second key feature and their SHAP values, respectively, to illustrate the degree of influence of a single feature on the model's prediction.

[0025] In addition, each of the second key features is further divided into a laboratory dataset, a treatment order dataset, a basic dataset, and other datasets according to the type. The laboratory dataset includes the latest blood glucose (Glu_latest). The dataset includes the latest serum potassium (K_latest), serum sodium (NA_latest), serum creatinine (CREAT_latest), latest white blood cell count (WBC_latest), latest neutrophil count (NEUT_latest), latest C-reactive protein (CRP_latest), latest blood urea nitrogen (BUN_latest), latest heme (Hgb_latest), latest glomerular filtration rate (EGFR_latest), and latest platelet count (PLT_latest). The treatment order dataset includes nasogastric tube (NG), urinary catheter (Foley), restraint (Restraint), and fasting (NPO). The basic dataset includes gender (Gender) and age (Age). Other datasets include hospitalizations (H_times), emergency room visits (ER_times), frailty index (FI), and emergency triage level (Triage).

[0026] The selection method for the features included in the second key feature was obtained through actual test results, namely the model performance evaluation procedure S3. In short, during the testing process, weakened classification models constructed with different numbers of features (i.e., 77, 72, 61, 56, 50, and 46 features) of the second key feature were trained, and the accuracy and area under the curve (AUROC) of these weakened classification models were compared. The test results showed that the weakened classification model constructed with the aforementioned 21 features was the best. For example, Figure 3 shows the evaluation results of the area under the curve. These features were selected from mean potassium (K_mean), K_variance, and maximum potassium (K_...). The following are the parameters for blood urea nitrogen (BUN) and blood urea nitrogen (BUN) in relation to mean NA (mean), minimum serum potassium (K_min), latest serum potassium (K_latest), mean serum sodium (NA_mean), NA_variance, maximum serum sodium (NA_max), minimum serum sodium (NA_min), latest serum sodium (NA_latest), mean serum creatinine (CREAT_mean), NA_variance, maximum serum creatinine (CREAT_max), minimum serum creatinine (CREAT_min), latest serum creatinine (CREAT_latest), mean white blood cell count (WBC_mean), NA_variance, maximum white blood cell count (WBC_max), minimum white blood cell count (WBC_min), latest white blood cell count (WBC_latest), mean blood urea nitrogen (BUN_mean), and NA_variance. The following are the parameters for blood urea nitrogen (BUN), blood urea nitrogen (BUN_max), blood urea nitrogen (BUN_min), blood urea nitrogen (BUN_latest), mean hemoglobin (HGB_mean), hemoglobin variation (HGB_variance), mean glomerular filtration rate (EGFR_mean), glomerular filtration rate variation (EGFR_variance), mean platelet count (PLT_mean), platelet variation (PLT_variance), maximum platelet count (PLT_max), minimum platelet count (PLT_min), and latest platelet count (PLT_latest).Mean blood glucose (GLU_mean), GLU_variance, maximum blood glucose (GLU_max), minimum blood glucose (GLU_min), latest blood glucose (GLU_latest), mean neutrophils (NEUT_mean), NEUT_variance, maximum neutrophils (NEUT_max), minimum neutrophils (NEUT_min), latest neutrophils (NEUT_latest), mean C-reactive protein (CRP_mean), CRP_variance, maximum C-reactive protein (CRP_max). x), minimum C-reactive protein (CRP_min), latest C-reactive protein (CRP_latest), nasogastric tube (NG), urinary catheter (Foley), restraint, fasting (NPO), gender, age, number of hospitalizations (H_times), number of emergency room visits (ER_times), total Charson comorbidity index (CCI_sum), total frailty index (FI_sum), frailty index (FI), minimum body mass index (BMI_min), variable body mass index (BMI_var), emergency triage level (Triage), body weight, antidepressants, first-generation antihistamines, anticholinergics, antiparkinsonian agents, antipsychotics, and antispasmodics. The different numbers of characteristics are based on these characteristics and are reduced accordingly. Among them, blood glucose, blood potassium, blood sodium, blood creatinine, white blood cells, neutrophils, C-reactive protein, blood urea nitrogen, heme, glomerular filtration rate, and platelets were data measured in the laboratory. In addition, recursive feature elimination with cross-validation (RFECV) was used for feature selection; in other words, cross-validation was used to retain the most important features. The selection of different numbers of features and the number of features retained after cross-validation are shown in Table 1 below.

[0027] In addition, various judgment indicators under different algorithms with different numbers of features were tested in practice, namely Sensitivity, Spec, Accuracy, AUROC, and Precision, as shown in Tables 1 to 7 below.

[0028] Table 1. Explanation of the number of different features quantity Screening using Recursive Feature Removal (RFECV) combined with cross-validation illustrate 77 All features 72 Remove the maximum, minimum, average, variable, and most recent Body Mass Index (BMI). 61 twenty one Laboratory data on removing variants 56 29 Remove data including emergency room visits, hospitalizations, emergency triage, triage level, total Charlesson comorbidity index, and total frailty index. 50 29 Remove medications (antidepressants, first-generation antihistamines, anticholinergics, anti-Parkinson's drugs, antipsychotics, anticonvulsants) 46 7 Remove nasogastric tube, urinary catheter, restraints, and fasting.

[0029] Table 2. Test results of 77 features with different algorithms Classifier Sens Spec Accuracy AUROC Xgboost 76.92% 69.61% 75.82% 81.05% RF 71.85% 73.53% 72.11% 81.46% LR 72.90% 77.45% 73.59% 81.59% SVM 77.10% 69.61% 75.96% 82.66% KNN 95.63% 18.63% 83.98% 69.37%

[0030] Table 3. Test results of 72 features with different algorithms Classifier Sens Spec Accuracy AUROC Xgboost 77.27% 70.59% 76.26% 81.24% RF 71.85% 78.43% 72.85% 81.90% LR 73.60% 77.45% 74.18% 81.48% SVM 77.27% 72.55% 76.56% 82.32% KNN 95.63% 17.65% 83.83% 66.62%

[0031] Table 4. Test results of 61 features and the number of 21 features selected by RFECV under different algorithms. Feature number Classifier Sens Spec Accuracy AUROC Precision 61 Xgboost 75.52% 71.57% 74.93% 80.75% 93.71% RF 71.50% 77.45% 72.40% 81.74% 94.68% LR 72.55% 76.47% 73.15% 81.58% 94.53% SVM 77.97% 71.57% 77.00% 82.67% 93.89% KNN 95.98% 22.55% 84.87% 68.11% 87.42% 21 Xgboost 74.48% 69.61% 73.74% 80.96% 93.22% RF 72.55% 76.47% 73.15% 80.99% 94.53% LR 71.50% 77.45% 72.40% 80.61% 94.68% SVM 74.13% 78.43% 74.78% 82.38% 95.07% KNN 95.63% 18.63% 83.98% 71.90% 86.83%

[0032] Table 5. Test results of 56 features and the number of 29 features selected by RFECV under different algorithms. Feature number Classifier Sens Spec Accuracy AUROC Precision 56 Xgboost 74.65% 69.61% 73.89% 80.13% 93.23% RF 71.50% 78.43% 72.55% 80.53% 94.90% LR 73.08% 72.55% 73.00% 81.45% 93.72% SVM 77.80% 68.63% 76.41% 81.70% 93.29% KNN 96.33% 15.69% 84.12% 67.58% 86.50% 29 Xgboost 74.48% 70.59% 73.89% 80.04% 93.42% RF 73.25% 79.41% 74.18% 80.93% 95.23% LR 71.15% 75.49% 71.81% 80.32% 94.21% SVM 74.83% 72.55% 74.48% 80.96% 93.86% KNN 96.15% 19.61% 84.57% 69.49% 87.03%

[0033] Table 6. Test results of 50 features and the number of 29 features selected by RFECV under different algorithms. Feature number Classifier Sens Spec Accuracy AUROC Precision 50 Xgboost 74.65% 68.63% 73.74% 79.96% 93.03% RF 71.85% 78.43% 72.85% 81.04% 94.92% LR 73.08% 73.53% 73.15% 81.11% 93.93% SVM 74.83% 71.57% 74.33% 81.08% 93.65% KNN 96.33% 16.67% 84.27% 69.66% 86.64% 29 Xgboost 74.83% 69.61% 74.04% 80.00% 93.25% RF 71.15% 78.43% 72.26% 81.03% 94.87% LR 70.98% 76.47% 71.81% 80.14% 94.42% SVM 74.13% 73.53% 74.04% 81.08% 94.01% KNN 95.28% 19.61% 83.83% 69.52% 86.92%

[0034] Table 7. Test results of 46 features and the number of 7 features selected by RFECV in different algorithms. Eigenvalues Classifier Sens Spec Accuracy AUROC Precision 46 Xgboost 73.60% 68.63% 72.85% 78.28% 92.94% RF 71.85% 75.49% 72.40% 79.81% 94.27% LR 71.50% 69.61% 71.22% 78.59% 92.95% SVM 74.65% 67.65% 73.59% 78.84% 92.83% KNN 95.45% 11.76% 82.79% 66.86% 85.85% 7 Xgboost 75.87% 68.63% 74.78% 77.84% 93.13% RF 72.03% 74.51% 72.40% 77.25% 94.06% LR 72.20% 73.53% 72.40% 77.78% 93.86% SVM 72.73% 74.51% 73.00% 78.13% 94.12% KNN 95.45% 17.65% 83.68% 66.47% 86.67%

[0035] Table 8. Test results of the number of 21 features for different algorithms Classifier Sens Spec Accuracy AUROC Precision Xgboost 73.95% 66.67% 72.85% 80.62% 92.56% RF 70.63% 76.47% 71.51% 81.27% 94.39% LR 72.73% 74.51% 73.00% 80.77% 94.12% SVM 74.83% 72.55% 74.48% 80.84% 93.86% KNN 94.58% 21.57% 83.53% 72.52% 87.12%

[0036] In the model performance evaluation program S3, the processing module 20 analyzes and evaluates the weakened classification model constructed by different algorithms. For example, based on the evaluation data of different algorithms in Tables 1 to 8, the best algorithm is determined, and random forest and support vector machine have achieved better prediction results.

[0037] In addition, in the model performance evaluation procedure S3, features are selected through different feature compositions and by using a recursive feature deletion method combined with cross-validation, for example:

[0038] 1. The model was not improved when comparing feature 77 with feature 72 (i.e., the body mass index feature was removed) (as shown in Tables 2 and 3).

[0039] 2. The model was not improved when 72 features were compared with 61 features (i.e., features related to laboratory data of variance (i.e., variance of serum potassium (K_variance), variance of serum potassium (K_variance), variance of serum creatinine (CREAT_variance), variance of white blood cell count (WBC_variance), variance of blood urea nitrogen (BUN_variance), variance of heme (HGB_variance), variance of glomerular filtration rate (EGFR_variance), variance of blood glucose (GLU_variance), variance of neutrophils (NEUT_variance), and variance of C-reactive protein (CRP_variance))) (as shown in Tables 3 and 4).

[0040] 3. The model was compared with the 50 features (i.e., 6 drug features were deleted), and there was no improvement (as shown in Tables 5 and 6).

[0041] 4. Comparing different features with the features selected by RFECV, the model results are similar (as shown in Tables 4 to 7).

[0042] Based on the above comparison results, the variance-related features of the experimental data and the drug had little impact on the model. Therefore, the variance-related features and drug features of the experimental data were deleted. Considering the convenience of clinical application, the most recent data from the laboratory was used, resulting in a total of 21 features: 11 features from the laboratory dataset, 4 features from the treatment order dataset, 2 features from the basic dataset, and 4 features from the other datasets. Furthermore, the results for different algorithms for these 21 features are shown in Table 8. For the 21 features, the sensitivity of Random Forest was 70.63%, the specificity was 76.47%, and the area under the ROC curve was 81.27%; the sensitivity of Support Vector Machine was 74.83%, the specificity was 72.55%, and the area under the ROC curve was 80.84%.

[0043] Next, the processing module 20 also uses the data cleaning procedure S2 to preprocess these second key features, such as estimating missing values ​​and dropping outliers. In this example, the data cleaning procedure S2 mainly determines whether the second key feature is abnormal, and deletes the second key feature and health status of the hospitalized patient to which the second key feature is determined to be abnormal from the training data of the frailty classification model, so as to make the trained frailty classification model more accurate. Furthermore, the data cleaning procedure S2 uses missing data, data excluding pre-discharge assessment, hospitalization of less than 3 days, and completion of comprehensive geriatric assessment more than four days after hospitalization as the basis for determining whether each second key feature is abnormal, as described later.

[0044] Regarding the state of 'missing data', firstly, the data cleaning procedure S2 causes the processing module 20 to check whether the latest blood glucose (Glu_latest), latest blood potassium (K_latest), latest blood sodium (NA_latest), latest blood creatinine (CREAT_latest), latest white blood cell (WBC_latest), latest neutrophil (NEUT_latest), latest C-reactive protein (CRP_latest), latest blood urea nitrogen (BUN_latest), latest heme (Hgb_latest), latest glomerular filtration rate (EGFR_latest), and latest platelet (PLT_latest) in the laboratory dataset of each second key feature are missing more than a predetermined number, and the second key feature missing more than the predetermined number is defined as abnormal. In this example, the predetermined number is 6, while in other possible implementation states, the predetermined number can be arbitrarily increased or decreased, or even the number of missing data in other datasets can be checked.

[0045] For the case of "hospitalization for less than 3 days", the data cleaning program S2 enables the processing module 20 to confirm whether the number of days of hospitalization for each hospitalized patient is less than the analysis time point, and defines the second key feature extracted from the second medical record of the hospitalized patient whose number of days of hospitalization is less than the analysis time point as abnormal.

[0046] For the case of "comprehensive geriatric assessment completed more than four days after hospitalization", the data cleaning procedure S2 causes the processing module 20 to check whether the assessment date of each hospitalized patient's health status is greater than the analysis time point, i.e., the assessment date is 4 or more, and defines the second key feature extracted from the second medical record of the hospitalized patient whose assessment date is greater than the analysis time point as abnormal.

[0047] In this frailty identification procedure, the processing module 20 uses an analysis time point as a benchmark, reads the first medical data of the patient to be analyzed whose record time is earlier than the analysis time point and closest to the analysis time point, extracts a first key feature from the selected first medical data, and then inputs the first key feature into the frailty classification model to analyze a frailty identification result for the patient to be analyzed. The analysis time point is N days after the patient to be analyzed is hospitalized, where N is a positive integer; in this example, N is 3. The first key characteristic includes gender, age, latest blood glucose (Glu_latest), latest blood potassium (K_latest), latest blood sodium (NA_latest), latest blood creatinine (CREAT_latest), latest white blood cell count (WBC_latest), latest neutrophil count (NEUT_latest), latest C-reactive protein (CRP_latest), latest blood urea nitrogen (BUN_latest), latest heme (Hgb_latest), latest glomerular filtration rate (EGFR_latest), latest platelet count (PLT_latest), nasogastric tube (NG), urinary catheter (Foley), restraint (Restraint), fasting (NPO), number of hospitalizations (H_times), number of emergency room visits (ER_times), frailty index (FI), and emergency triage level (Triage).

[0048] Accordingly, medical staff can take appropriate medical measures based on the frailty identification results. For example, if the patient to be analyzed is confirmed to be frail, early intervention care can be provided, which can help reduce the medical burden caused by the patient's frailty, avoid inappropriate drug use, and reduce hospitalization time and mortality risk.

[0049] The above-described embodiments are only for illustrating the present invention and are not intended to limit the scope of the present invention. All modifications or variations made without departing from the spirit of the present invention are within the scope of protection intended by the present invention. [Simplified Explanation of the Diagram]

[0016] Figure 1 is a system block diagram of the inpatient frailty identification system according to a preferred embodiment of the present invention. Figure 2 is a SHAP diagram of each feature in the second key feature. Figure 3 is an analysis diagram of the area under the curve. Figures 4 to 24 are relationship diagrams obtained by different features and their SHAP values. Figure 25 is a flowchart of a preferred embodiment of the present invention.

Claims

1. A frailty identification system for hospitalized patients, comprising: A database stores the first medical record of a patient to be analyzed and a frailty classification model. The first medical record contains multiple sets of first medical data recorded at different time points during the patient's hospitalization. A processing module, using an analysis time point as a reference, reads the first medical data of the patient's first medical record that is earlier than and closest to the analysis time point. It extracts a first key feature from the selected first medical data and inputs this first key feature into the frailty classification model to analyze and generate a frailty identification result for the patient. The analysis time point is N days after the patient's hospitalization, where N is a positive integer. The first key feature includes gender, age, and latest blood glucose level (Glu_latest). The database contains the following data: latest serum potassium (K_latest), latest serum sodium (NA_latest), latest serum creatinine (CREAT_latest), latest white blood cell count (WBC_latest), latest neutrophil count (NEUT_latest), latest C-reactive protein (CRP_latest), latest blood urea nitrogen (BUN_latest), latest heme (Hgb_latest), latest glomerular filtration rate (EGFR_latest), latest platelet count (PLT_latest), nasogastric tube (NG), urinary catheter (Foley), restraint (Restraint), fasting (NPO), number of hospitalizations (H_times), number of emergency room visits (ER_times), frailty index (FI), and emergency triage level (Triage). The database also stores multiple second medical records for different hospitalized patients. Each second medical record contains multiple second medical data collected at different time points during the patient's hospitalization and the patient's health status, which is either non-frail or frail. The processing module reads the second medical data recorded last in the second medical record according to the time sequence, and extracts a second key feature from the selected second medical data. Then, the second key features and the health status of the hospitalized patients are used as training data to construct the frailty classification model.The second key feature was selected through a model performance evaluation procedure. The second key feature includes gender, age, latest blood glucose (Glu_latest), latest blood potassium (K_latest), latest blood sodium (NA_latest), latest blood creatinine (CREAT_latest), latest white blood cell count (WBC_latest), latest neutrophil count (NEUT_latest), latest C-reactive protein (CRP_latest), latest blood urea nitrogen (BUN_latest), latest heme (Hgb_latest), latest glomerular filtration rate (EGFR_latest), latest platelet count (PLT_latest), nasogastric tube (NG), urinary catheter (Foley), restraint (Restraint), fasting (NPO), number of hospitalizations (H_times), number of emergency room visits (ER_times), frailty index (FI), and emergency triage level (Triage). The model performance evaluation procedure involves training and evaluating weakened classification models constructed using the second key feature with varying numbers of features, and selecting the second key feature corresponding to the weakened classification model with the best performance.

2. The inpatient frailty identification system as described in claim 1, wherein, In the model performance evaluation procedure, the second key feature is also selected using recursive feature elimination with cross-validation (RFECV).

3. The inpatient frailty identification system as described in claim 2, wherein, Each of these second key features is further categorized into a laboratory dataset, a treatment order dataset, a basic dataset, and other datasets. The laboratory dataset includes the latest blood glucose (Glu_latest), latest blood potassium (K_latest), latest blood sodium (NA_latest), latest blood creatinine (CREAT_latest), latest white blood cell count (WBC_latest), latest neutrophil count (NEUT_latest), latest C-reactive protein (CRP_latest), latest blood urea nitrogen (BUN_latest), latest heme (Hgb_latest), latest glomerular filtration rate (EGFR_latest), and latest platelet count (PLT_latest). The treatment order dataset includes nasogastric tube (NG), urinary catheter (Foley), restraint (Restraint), and fasting (NPO). The basic dataset includes gender (Gender) and age (Age). The other dataset includes hospitalizations (H_times), emergency room visits (ER_times), frailty index (FI), and emergency triage level (Triage).

4. The inpatient frailty identification system as described in claim 3, wherein, The processing module also uses a data cleaning procedure to determine whether the second key feature is abnormal, and deletes the second key feature and health status of the hospitalized patient to which the second key feature is determined to be abnormal from the training data of the frailty classification model.

5. The inpatient frailty identification system as described in claim 4, wherein, The data cleaning procedure refers to the processing module checking whether the latest blood glucose (Glu_latest), latest blood potassium (K_latest), latest blood sodium (NA_latest), latest blood creatinine (CREAT_latest), latest white blood cell (WBC_latest), latest neutrophil (NEUT_latest), latest C-reactive protein (CRP_latest), latest blood urea nitrogen (BUN_latest), latest heme (Hgb_latest), latest glomerular filtration rate (EGFR_latest), and latest platelet (PLT_latest) in the laboratory dataset for each of the second key features is missing more than a predetermined number, and defining the second key feature missing more than the predetermined number as abnormal.

6. The inpatient frailty identification system as described in claim 5, wherein, The pre-order quantity is 6.

7. The inpatient frailty identification system as described in claim 4, wherein, The data cleaning procedure refers to the processing module confirming whether the length of hospital stay for each hospitalized patient is less than the analysis time point, and defining the second key feature extracted from the second medical record of hospitalized patients whose length of hospital stay is less than the analysis time point as an anomaly.

8. The inpatient frailty identification system as described in claim 4, wherein, The data cleaning procedure involves the processing module checking whether the date on which each hospitalized patient completed the health status assessment is greater than the analysis time point, and defining the second key feature extracted from the second medical record of hospitalized patients whose assessment date is greater than the analysis time point as an anomaly.

9. The inpatient frailty identification system as described in claim 1, wherein, N is 3.

10. The inpatient frailty identification system as described in claim 1, wherein, The weakened classification model is a group consisting of Extreme Gradient Boosting (XGBoost), Support Vector Machine (SVM), k-nearest neighbor (KNN), Random Forest (RF) algorithm, and Logistic Regression (LR).

Images