Objective To summarize advances in the application of machine learning in the diagnosis and treatment of liver disease. Method The recent literatures on the progress of machine learning in the diagnosis, treatment and prognosis of liver diseases were reviewed. Results Machine learning could be used to diagnose and categorize substantial liver lesions, tumourous lesions and rare liver diseases at an early stage, which could facilitate clinicians to take timely and appropriate treatment measures. Machine learning was helpful in informing clinicians in choosing the best treatment decision, which was conducive to reducing medical risks. It could also help to determine the prognosis of patients in a comprehensive manner, and provide assistance in formulating early rehabilitation treatment plans, adjusting follow-up strategies and improving future prognosis. Conclusions Multiple types of machine learning algorithms have achieved positive results in the clinical application of liver diseases by constructing different prediction models, and have great potential and excellent prospects in multiple aspects such as diagnosis, treatment and prognosis of liver diseases.
Inadvertent perioperative hypothermia (IPH) is one of the common complications of surgery, which can lead to a series of adverse consequences. In recent years, with the deepening development of precision medicine concepts, establishing predictive models to identify the risk of IPH early and implementing targeted interventions has become an important research direction for perioperative management. This article reviews the current research status of IPH predictive models in adults, focusing on the research design, modeling methods, selection of prediction factors, and prediction performance of different predictive models. It also explores the advantages and limitations of existing models, aiming to provide references for the selection, application, and optimization of relevant predictive models.
ObjectiveTo analyze the relevant risk factors affecting postoperative relapse-free survival (RFS) in the primary gastrointestinal stromal tumors (GIST) and develop a Nomogram predictive model of postoperative RFS for the GIST patients. MethodsThe patients diagnosed with GIST by postoperative pathology from January 2011 to December 2020 at the First Hospital of Lanzhou University and Gansu Provincial People’s Hospital were collected, and then were randomly divided into a training set and a validation set at a ratio of 7∶3 using R software function. The univariate and multivariate Cox regression analysis were used to identify the risk factors affecting the RFS for the GIST patients after surgery, and then based on this, the Nomogram predictive model was constructed to predict the probability of RFS at 3- and 5-year after surgery for the patients with GIST. The effectiveness of the Nomogram was evaluated using the area under the receiver operating characteristic curve (AUC), consistency index (C-index), and calibration curve, and the clinical utility of the Nomogram and the modified National Institutes of Health (M-NIH) classification standard was evaluated using the decision curve analysis (DCA). ResultsA total of 454 patients were included, including 317 in the training set and 137 in the validation set. The results of multivariate Cox regression analysis showed that the tumor location, tumor size, differentiation degree, American Joint Committee onCancer TNM stage, mitotic rate, CD34 expression, treatment method, number of lymph node detection, and targeted drug treatment time were the influencing factors of postoperative RFS for the GIST patients (P<0.05). The Nomogram predictive model was constructed based on the influencing factors. The C-index of the Nomogram in the training set and validation set were 0.731 [95%CI (0.679, 0.783)] and 0.685 [95%CI (0.647, 0.722)], respectively. The AUC (95%CI) of distinguishing the RFS at 3- and 5-year after surgery were 0.764 (0.681, 0.846) and 0.724 (0.661, 0.787) in the training set and 0.749 (0.625, 0.872) and 0.739 (0.647, 0.832) in the validation set, respectively. The calibration curve results showed that a good consistency of the 3-year and 5-year recurrence free survival rates between the predicted results and the actual results in the training set, while which was slightly poor in the validation set. There was a higher net benefit for the 3-year recurrence free survival rate after GIST surgery when the threshold probability range was 0.19 to 0.57. When the threshold probability range was 0.44 to 0.83, there was a higher net benefit for the 5-year recurrence free survival rate after GIST surgery. And within the threshold probability ranges, the net benefit of the Nomogram was better than the M-NIH classification system at the corresponding threshold probability. ConclusionsThe results of this study suggest that the patients with GIST located in the other sites (mainly including the esophagus, duodenum, and retroperitoneum), with tumor size greater than 5 cm, poor or undifferentiated differentiation, mitotic rate lower than 5/50 HPF, negative CD34 expression, ablation treatment, number of lymph nodes detected more than 4, and targeted drug treatment time less than 3 months need to closely pay attentions to the postoperative recurrence. The discrimination and clinical applicability of the Nomogram predictive model are good.
Objective To explore the independent risk factors for hospital infections in tertiary hospitals in Gansu Province, and establish and validate a prediction model. Methods A total of 690 patients hospitalized with hospital infections in Gansu Provincial Hospital between January and December 2021 were selected as the infection group; matched with admission department and age at a 1∶1 ratio, 690 patients who were hospitalized during the same period without hospital infections were selected as the control group. The information including underlying diseases, endoscopic operations, blood transfusion and immunosuppressant use of the two groups were compared, the factors influencing hospital infections in hospitalized patients were analyzed through multiple logistic regression, and the logistic prediction model was established. Eighty percent of the data from Gansu Provincial Hospital were used as the training set of the model, and the remaining 20% were used as the test set for internal validation. Case data from other three hospitals in Gansu Province were used for external validation. Sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve (AUC) were used to evaluate the model effectiveness. Results Multiple logistic regression analysis showed that endoscopic therapeutic manipulation [odds ratio (OR)=3.360, 95% confidence interval (CI) (2.496, 4.523)], indwelling catheter [OR=3.100, 95%CI (2.352, 4.085)], organ transplantation/artifact implantation [OR=3.133, 95%CI (1.780, 5.516)], blood or blood product transfusions [OR=3.412, 95%CI (2.626, 4.434)], glucocorticoids [OR=2.253, 95%CI (1.608, 3.157)], the number of underlying diseases [OR=1.197, 95%CI (1.068, 1.342)], and the number of surgical procedures performed during hospitalization [OR=1.221, 95%CI (1.096, 1.361)] were risk factors for hospital infections. The regression equation of the prediction model was: logit(P)=–2.208+1.212×endoscopic therapeutic operations+1.131×indwelling urinary catheters+1.142×organ transplantation/artifact implantation+1.227×transfusion of blood or blood products+0.812×glucocorticosteroids+0.180×number of underlying diseases+0.200×number of surgical procedures performed during the hospitalization. The internal validation set model had a sensitivity of 72.857%, a specificity of 77.206%, an accuracy of 76.692%, and an AUC value of 0.817. The external validation model had a sensitivity of 63.705%, a specificity of 70.934%, an accuracy of 68.669%, and an AUC value of 0.726. Conclusions Endoscopic treatment operation, indwelling catheter, organ transplantation/artifact implantation, blood or blood product transfusion, glucocorticoid, number of underlying diseases, and number of surgical cases during hospitalization are influencing factors of hospital infections. The model can effectively predict the occurrence of hospital infections and guide the clinic to take preventive measures to reduce the occurrence of hospital infections.
ObjectiveTo identify the risk factors of bone metastasis in breast cancer and construct a predictive model. MethodsThe data of breast cancer patients met inclusion and exclusion criteria from 2010 to 2015 were obtained from the SEER*Stat database. Additionally, the data of breast cancer patients diagnosed with distant metastasis in the Affiliated Hospital of Southwest Medical University from 2021 to 2023 were collected. The patients from the SEER database were randomly divided into training (70%) and validation (30%) sets using R software, and the breast cancer patients from the Affiliated Hospital of Southwest Medical University were included in the validation set. The univariate and multivariate logistic regressions were used to identify risk factors of breast cancer bone metastasis. A nomogram predictive model was then constructed based on these factors. The predictive effect of the nomogram was evaluated using the area under the receiver operating characteristic curve (AUC), calibration curve, and decision curve analysis. ResultsThe study included 8 637 breast cancer patients, with 5 998 in the training set and 2 639 (including 68 patients in the Affiliated Hospital of Southwest Medical University) in the validation set. The statistical differences in the race and N stage were observed between the training and validation sets (P<0.05). The multivariate logistic regression analysis revealed that being of white race, having a low histological grade (Ⅰ–Ⅱ), positive estrogen and progesterone receptors status, negative human epidermal growth factor receptor 2 status, and non-undergoing surgery for the primary breast cancer site increased the risk of breast cancer bone metastasis (P<0.05). The nomogram based on these risk factors showed that the AUC (95% CI) of the training and validation sets was 0.676 (0.533, 0.744) and 0.690 (0.549, 0.739), respectively. The internal calibration using 1 000 Bootstrap samples demonstrated that the calibration curves for both sets closely approximated the ideal 45-degree reference line. The decision curve analysis indicated a stronger clinical utility within a certain probability threshold range. ConclusionsThis study constructs a nomogram predictive model based on factors related to the risk of breast cancer bone metastasis, which demonstrates a good consistency between actual and predicted outcomes in both training and validation sets. The nomogram shows a stronger clinical utility, but further analysis is needed to understand the reasons of the lower differentiation of nomogram in both sets.
ObjectiveTo construct a new model for predicting the overall survival rate of gastric cancer and to guide the clinical work.MethodsThe clinical information and gene expression information of patients with gastric cancer were downloaded through The Cancer Genome Atlas (TCGA) database. The clinicopathologic characteristics and gene expression information affecting the overall survival rate of gastric cancer patients were screened by univariate COX regression and Lasson regression, then the predictive model was constructed by multiple COX regression model, and the predictive model was tested by receiver operating characteristic curve, calibration curve and decision curve analysis curve. The effect of genes included in the predictive model on the overall survival rate of patients with gastric cancer was discussed, and the predictive model diagram was drawn.ResultsThrough repeated screening and comparison of the model, the patient’s age, T stage, N stage, M stage and 12 genes (INCENP, IGHD3-16, ITFG1-AS1, NEK5, MATN3, YWHABP2, SYT12, LINC01210, ZNF385C, LINC01980, CYMP-AS1 and FAT3) were included in the predictive model. The prediction ability of this model was close to or more than 80%, which was significantly higher than that of the traditional TNM staging prediction system. All the indexes included in the model were significantly different by univariate and multivariate COX regression analysis（P<0.05）, and the 12 genes included were the risk factors affecting the overall survival rate of gastric cancer.ConclusionThe gastric cancer prediction model constructed by combining clinical characteristics and genomics has good predictive ability and can guide clinical work.
ObjectiveTo analyze the prognostic factors of patients with bacterial bloodstream infection sepsis and to identify independent risk factors related to death, so as to potentially develop one predictive model for clinical practice. Method A non-intervention retrospective study was carried out. The relative data of adult sepsis patients with positive bacterial blood culture (including central venous catheter tip culture) within 48 hours after admission were collected from the electronic medical database of the First Affiliated Hospital of Dalian Medical University from January 1, 2018 to December 31, 2019, including demographic characters, vital signs, laboratory data, etc. The patients were divided into a survival group and a death group according to in-hospital outcome. The risk factors were analyzed and the prediction model was established by means of multi-factor logistics regression. The discriminatory ability of the model was shown by area under the receiver operating characteristic curve (AUC). The visualization of the predictive model was drawn by nomogram and the model was also verified by internal validation methods with R language. Results A total of 1189 patients were retrieved, and 563 qualified patients were included in the study, including 398 in the survival group and 165 in the death group. Except gender and pathogen type, other indicators yielded statistical differences in single factor comparison between the survival group and the death group. Independent risk factors included in the logistic regression prediction model were: age [P=0.000, 95% confidence interval (CI) 0.949 - 0.982], heart rate (P=0.000, 95%CI 0.966 - 0.987), platelet count (P=0.009, 95%CI 1.001 - 1.006), fibrinogen (P=0.036, 95%CI 1.010 - 1.325), serum potassium ion (P=0.005, 95%CI 0.426 - 0.861), serum chloride ion (P=0.054, 95%CI 0.939 - 1.001), aspartate aminotransferase (P=0.03, 95%CI 0.996 - 1.000), serum globulin (P=0.025, 95%CI 1.006 - 1.086), and mean arterial pressure (P=0.250, 95%CI 0.995 - 1.021). The AUC of the prediction model was 0.779 (95%CI 0.737 - 0.821). The prediction efficiency of the total score of the model's nomogram was good in the 210 - 320 interval, and mean absolute error was 0.011, mean squared error was 0.00018. Conclusions The basic vital signs within 48 h admitting into hospital, as well those homeostasis disordering index indicated by coagulation, liver and renal dysfunction are highly correlated with the prognosis of septic patients with bacterial bloodstream infection. Early warning should be set in order to achieve early detection and rescue patients’ lives.
Objective To clarify the specific clinical predictive efficacy of CT and serological indicators for the progression of connective tissue disease-associated interstitial lung disease (CTD-ILD) to progressive pulmonary fibrosis (PPF). Methods Patients who were diagnosed with CTD-ILD in Chest Hospital of Zhengzhou University Between January 2020 and December 2021 were recruited in the study. Clinical data and high-resolution CT results of the patients were collected. The patients were divided into a stable group and a progressive group (PPF group) based on whether PPF occurred during follow-up. COX proportional hazards regression was used to identify risk factors affecting the progression of CTD-ILD to PPF, and a risk prediction model was established based on the results of the COX regression model. The predictive efficacy of the model was evaluated through internal cross-validation. Results A total of 194 patients diagnosed with CTD-ILD were enrolled based on the inclusion and exclusion criteria. Among them, 34 patients progressed to PPF during treatment, and 160 patients did not progress. The variables obtained at lambda$1se in LASSO regression were ANCA associated vasculitis, lymphocytes, albumin, erythrocyte sedimentation rate, and serum ferritin. Multivariate COX regression analysis showed that the extent of fibrosis, serum ferritin, albumin, and age were independent risk factors for the progression of CTD-ILD to PPF (all P<0.05). A prediction model was established based on the results of the multivariate COX regression analysis. The area under the receiver operator characteristic curve at 6 months, 9 months, and 12 months was 0.989, 0.931, and 0.797, respectively, indicating that the model has good discrimination and sensitivity, and good predictive efficacy. The calibration curve showed a good overlap between predicted and actual values. Conclusions The extent of fibrosis, serum ferritin, albumin, and age are independent risk factors for the progression of CTD-ILD to PPF. The model established based on this and externally validated shows good predictive efficacy.
ObjectiveTo explore the risk factors which affect the severity of acute pancreatitis by using machine learning algorithms. MethodsA retrospective review was conducted of medical records from 262 patients hospitalized for acute pancreatitis at the Second Affiliated Hospital of Zhengzhou University between October 2022 and February 2024. Patients were classified according to the revised edition Atlanta Classification into mild cases (n=146) and non-mild cases (n=116). LASSO analysis was employed to identify predictors for non-mild acute pancreatitis. Six machine learning algorithms, including extreme gradient boosting, random forest, logistic regression, decision tree, support vector machine, and K-nearest neighbors were integrated to construct predictive models. Model performance was evaluated by comparing the following metrics: area under the curve (AUC), sensitivity, specificity, accuracy, F1 score, calibration curves, and decision curves. ResultsThrough LASSO regression analysis, six feature variables, including heart rate, white blood cell count, neutrophil count, C-reactive protein, albumin, and calcium ion were selected to train and test machine learning models. Results showed that extreme gradient boosting achieved the highest AUC value of 0.93 on the test set, making it the optimal model. The sensitivity, specificity, accuracy, Brier score, and F1 score of the extreme gradient boosting model were 0.97, 0.70, 0.85, 0.108, and 0.84. ConclusionThe prediction model developed using extreme gradient boosting has high clinical utility value, helps to predict the severity of acute pancreatitis at an early stage and is valuable in guiding clinical decision-making.
ObjectiveTo explore the predictive value of a simple prediction model for patients with acute myocardial infarction.MethodsClinical data of 280 patients with acute ST-segment elevation myocardial infarction (STEMI) in the Department of Emergence Medicine, West China Hospital of Sichuan University from January 2019 to January 2020 were retrospectively analyzed. The patients were divided into a death group (n=34) and a survival group (n=246).ResultsAge, heart rate, body mass index (BMI), global registry of acute coronary events (GRACE), thrombolysis in myocardial infarction trial (TIMI) score, blood urea nitrogen, serum cystatin C and D-dimer in the survival group were less or lower than those in the death group (P<0.05). Left ventricle ejection fraction and the level of albumin, triglyceride, total cholesterol and low density lipoprotein cholesterol were higher and the incidence of Killip class≥Ⅲ was lower in the survival group compared to the death group (P<0.05). Multivariate logistic regression analysis showed that age, BMI, heart rate, diastolic blood pressure, and systolic blood pressure were independent risk factors for all-cause death in STEMI patients. Receiver operating characteristic (ROC) curve analysis showed that the area under the curve of simple prediction model for predicting death was 0.802, and similar to that of GRACE (0.816). The H-L test showed that the simple model had high accuracy in predicting death (χ2=3.77, P=0.877). Pearson correlation analysis showed that the simple prediction model was significantly correlated with the GRACE (r=0.651, P<0.001) and coronary artery stenosis score (r=0.210, P=0.001).ConclusionThe simple prediction model may be used to predict the hospitalization and long-term outcomes of STEMI patients, which is helpful to stratify high risk patients and to guide treatment.