Application of interpretable machine learning algorithms to predict distant metastasis in ovarian clear cell carcinoma

Background Predicting the prognosis of patients with acute myocardial infarction (AMI) combined with diabetes mellitus (DM) is crucial due to high in-hospital mortality rates. This study aims to develop and validate a mortality risk prediction model for these patients by interpretable machine learning (ML) methods. Methods Data were sourced from the Medical Information Mart for Intensive Care IV (MIMIC-IV, version 2.2). Predictors were selected by Least absolute shrinkage and selection operator (LASSO) regression and checked for multicollinearity with Spearman’s correlation. Patients were randomly assigned to training and validation sets in an 8:2 ratio. Seven ML algorithms were used to construct models in the training set. Model performance was evaluated in the validation set using metrics such as area under the curve (AUC) with 95% confidence interval (CI), calibration curves, precision, recall, F1 score, accuracy, negative predictive value (NPV), and positive predictive value (PPV). The significance of differences in predictive performance among models was assessed utilising the permutation test, and 10-fold cross-validation further validated the model’s performance. SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) were applied to interpret the models. Results The study included 2,828 patients with AMI combined with DM. Nineteen predictors were identified through LASSO regression and Spearman’s correlation. The Random Forest (RF) model was demonstrated the best performance, with an AUC of 0.823 (95% CI: 0.774–0.872), high precision (0.867), accuracy (0.873), and PPV (0.867). The RF model showed significant differences (p < 0.05) compared to the K-Nearest Neighbours and Decision Tree models. Calibration curves indicated that the RF model’s predicted risk aligned well with actual outcomes. 10-fold cross-validation confirmed the superior performance of RF model, with an average AUC of 0.828 (95% CI: 0.800–0.842). Significant Variables in RF model indicated that the top eight significant predictors were urine output, maximum anion gap, maximum urea nitrogen, age, minimum pH, maximum international normalised ratio (INR), mean respiratory rate, and mean systolic blood pressure. Conclusion This study demonstrates the potential of ML methods, particularly the RF model, in predicting in-hospital mortality risk for AMI patients with DM. The SHAP and LIME methods enhance the interpretability of ML models.

Treatment with anti-HER2 therapy could be beneficial for patients with HER2-positive endometrial and ovarian clear cell carcinoma (CCC). We studied HER2 overexpression by immunohistochemistry (IHC) using three different antibodies, including concordance with amplification by in-situ hybridisation (ISH). IHC and ISH were performed on tissue microarrays of 101 tumours: 58 endometrial pure CCC, 19 endometrial mixed carcinomas with a CCC component and 24 ovarian pure CCC. IHC was performed using SP3, 4B5 and HercepTest antibodies, and was scored by two independent observers. ISH was performed using dual-colour silver ISH. Using IHC, agreement was poor between SP3/4B5 (61.4%), poor between SP3/HercepTest (68.3%) and reasonable between 4B5/HercepTest (75.2%). Interobserver agreement was substantial to almost perfect for all antibodies (SP3: linear weighted κ = 0.89, 4B5: κ = 0.90, HercepTest: κ = 0.76). HER2-positivity by ISH was 17.8% (endometrial pure CCC: 24.1%, endometrial mixed: 0%, ovarian pure CCC: 16.7%). IHC/ISH concordance was poor, with a high false-negative rate of all three IHC antibodies: sensitivity (38.9-50.0%) and positive predictive value (PPV) (37.5-58.3%) were poor; specificity (81.9-94.0%) and negative predictive value (NPV) (87.1-88.3%) were reasonable. When excluding 2+ cases, sensitivity declined (26.7-43.8%) but PPV (80.0-87.5%) and specificity (98.6-98.7%) improved. In ovarian and endometrial CCC, there is considerable difference in HER2 overexpression by different IHC antibodies and marked discordance with ISH. As such, no single antibody can be considered conclusive for determining HER2 status in CCC. Based on these results, the lack of predictive value of different HER2 testing methods, as used in other studies, could be explained.

Background and aimsThe birth weight of a newborn is a crucial factor that affects their overall health and future well-being. Low birth weight (LBW) is a widespread global issue, which the World Health Organization defines as weighing less than 2,500 g. LBW can have severe negative consequences on an individual’s health, including neonatal mortality and various health concerns throughout their life. To address this problem, this study has been conducted using BDHS 2017–2018 data to uncover important aspects of LBW using a variety of machine learning (ML) approaches and to determine the best feature selection technique and best predictive ML model.MethodsTo pick out the key features, the Boruta algorithm and wrapper method were used. Logistic Regression (LR) used as traditional method and several machine learning classifiers were then used, including, DT (Decision Tree), SVM (Support Vector Machine), NB (Naïve Bayes), RF (Random Forest), XGBoost (eXtreme Gradient Boosting), and AdaBoost (Adaptive Boosting), to determine the best model for predicting LBW. The model’s performance was evaluated based on the specificity, sensitivity, accuracy, F1 score and AUC value.ResultsResult shows, Boruta algorithm identifies eleven significant features including respondent’s age, highest education level, educational attainment, wealth index, age at first birth, weight, height, BMI, age at first sexual intercourse, birth order number, and whether the child is a twin. Incorporating Boruta algorithm’s significant features, the performance of traditional LR and ML methods including DT, SVM, NB, RF, XGBoost, and AB were evaluated where LR, had a specificity, sensitivity, accuracy and F1 score of 0.85, 0.5, 85.15% and 0.915. While the ML methods DT, SVM, NB, RF, XGBoost, and AB model’s respective accuracy values were 85.35%, 85.15%, 84.54%, 81.18%, and 84.41%. Based on the specificity, sensitivity, accuracy, F1 score and AUC, RF (specificity = 0.99, sensitivity = 0.58, accuracy = 85.86%, F1 score = 0.9243, AUC = 0.549) outperformed the other methods. Both the classical (LR) and machine learning (ML) models’ performance has improved dramatically when important characteristics are extracted using the wrapper method. The LR method identified five significant features with a specificity, sensitivity, accuracy and F1 score of 0.87, 0.33, 87.12% and 0.9309. The region, whether the infant is a twin, and cesarean delivery were the three key features discovered by the DT and RF models, which were implemented using the wrapper technique. All three models had the identical F1 score of 0.9318. However, “child is twin” was recognized as a significant feature by the SVM, NB, and AB models, with an F1 score of 0.9315. Ultimately, with an F1 score of 0.9315, the XGBoost model recognized “child is twin” and “age at first sex” as relevant features. Random Forest again beat the other approaches in this instance.ConclusionsThe study reveals Wrapper method as the optimal feature selection technique. The ML method outperforms traditional methods, with Random Forest (RF) being the most effective predictive model for Low-Birth-Weight prediction. The study suggests that policymakers in Bangladesh can mitigate low birth weight newborns by considering identified risk factors.

Although the goal of rectal cancer treatment is to restore gastrointestinal continuity, some patients with rectal cancer develop a permanent stoma (PS) after sphincter-saving operations. Although many studies have identified the risk factors and causes of PS, few have precisely predicted the probability of PS formation before surgery. To validate whether an artificial intelligence model can accurately predict PS formation in patients with rectal cancer after sphincter-saving operations. Patients with rectal cancer who underwent a sphincter-saving operation at Taipei Medical University Hospital between January 1, 2012, and December 31, 2021, were retrospectively included in this study. A machine learning technique was used to predict whether a PS would form after a sphincter-saving operation. We included 19 routinely available preoperative variables in the artificial intelligence analysis. To evaluate the efficiency of the model, 6 performance metrics were utilized: accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and area under the receiving operating characteristic curve. In our classification pipeline, the data were randomly divided into a training set (80% of the data) and a validation set (20% of the data). The artificial intelligence models were trained using the training dataset, and their performance was evaluated using the validation dataset. Synthetic minority oversampling was used to solve the data imbalance. A total of 428 patients were included, and the PS rate was 13.6% (58/428) in the training set. The logistic regression (LR), Gaussian Naïve Bayes (GNB), Extreme Gradient Boosting (XGB), Gradient Boosting (GB), random forest, decision tree and light gradient boosting machine (LightGBM) algorithms were employed. The accuracies of the logistic regression (LR), Gaussian Naïve Bayes (GNB), Extreme Gradient Boosting (XGB), Gradient Boosting (GB), random forest (RF), decision tree (DT) and light gradient boosting machine (LightGBM) models were 70%, 76%, 89%, 93%, 95%, 79% and 93%, respectively. The area under the receiving operating characteristic curve values were 0.79 for the LR model, 0.84 for the GNB, 0.95 for the XGB, 0.95 for the GB, 0.99 for the RF model, 0.79 for the DT model and 0.98 for the LightGBM model. The key predictors that were identified were the distance of the lesion from the anal verge, clinical N stage, age, sex, American Society of Anesthesiologists score, and preoperative albumin and carcinoembryonic antigen levels. Integration of artificial intelligence with available preoperative data can potentially predict stoma outcomes after sphincter-saving operations. Our model exhibited excellent predictive ability and can improve the process of obtaining informed consent.

Introduction: Heart failure (HF) may induce bowel hypoperfusion, leading to hypoxia of the villa of the bowel wall and the occurrence of Clostridioides difficile infection (CDI). However, the risk factors for the development of CDI in HF patients have yet to be fully illustrated, especially because of a lack of evidence from real-world data. Methods: Clinical data and survival situations of HF patients with CDI admitted to ICU were extracted from the Medical Information Mart for Intensive Care (MIMIC)-IV database. For developing a model that can predict 28-day all-cause mortality in HF patients with CDI, the Recursive Feature Elimination with Cross-Validation (RFE-CV) method was used for feature selection. And nine machine learning (ML) algorithms, including logistic regression (LR), decision tree, Bayesian, adaptive boosting, random forest (RF), gradient boosting decision tree, XGBoost, light gradient boosting machine, and categorical boosting, were applied for model construction. After training and hyperparameter optimization of the models through grid search 5-fold cross-validation, the performance of models was evaluated by the area under curve (AUC), accuracy, sensitivity, specificity, precision, negative predictive value, and F1 score. Furthermore, the SHapley Additive exPlanations (SHAP) method was used to interpret the optimal model. Results: A total of 526 HF patients with CDI were included in the study, of whom 99 cases (18.8%) experienced death within 28 days. Eighteen of the 57 variables were selected for the model construction algorithm for model construction. Among the ML models considered, the RF model emerged as the optimal model achieving the accuracy, F1-score, and AUC values of 0.821, 0.596, and 0.864, respectively. The net benefit of the model surpassed other models at 16%–22% threshold probabilities based on decision curve analysis. According to the importance of features in the RF model, red blood cell distribution width, blood urea nitrogen, Simplified Acute Physiology Score II, Sequential Organ Failure Assessment, and white blood cell count were highlighted as the five most influential variables. Conclusions: We developed ML models to predict 28-day all-cause mortality in HF patients associated with CDI in the ICU, which are more effective than the conventional LR model. The RF model has the best performance among all the ML models employed. It may be useful to help clinicians identify high-risk HF patients with CDI.

Prediction of non-suicidal self-injury in adolescents at the family level using regression methods and machine learning

To predict the potential geographic distribution of Oncomelania hupensis in Yunnan Province using random forest (RF) and maximum entropy (MaxEnt) models, so as to provide insights into O. hupensis surveillance and control in Yunnan Province. The O. hupensis snail survey data in Yunnan Province from 2015 to 2016 were collected and converted into O. hupensis snail distribution site data. Data of 22 environmental variables in Yunnan Province were collected, including twelve climate variables (annual potential evapotranspiration, annual mean ground surface temperature, annual precipitation, annual mean air pressure, annual mean relative humidity, annual sunshine duration, annual mean air temperature, annual mean wind speed, ≥ 0 ℃ annual accumulated temperature, ≥ 10 ℃ annual accumulated temperature, aridity and index of moisture), eight geographical variables (normalized difference vegetation index, landform type, land use type, altitude, soil type, soil textureclay content, soil texture-sand content and soil texture-silt content) and two population and economic variables (gross domestic product and population). Variables were screened with Pearson correlation test and variance inflation factor (VIF) test. The RF and MaxEnt models and the ensemble model were created using the biomod2 package of the software R 4.2.1, and the potential distribution of O. hupensis snails after 2016 was predicted in Yunnan Province. The predictive effects of models were evaluated through cross-validation and independent tests, and the area under the receiver operating characteristic curve (AUC), true skill statistics (TSS) and Kappa statistics were used for model evaluation. In addition, the importance of environmental variables was analyzed, the contribution of environmental variables output by the models with AUC values of > 0.950 and TSS values of > 0.850 were selected for normalization processing, and the importance percentage of environmental variables was obtained to analyze the importance of environmental variables. Data of 148 O. hupensis snail distribution sites and 15 environmental variables were included in training sets of RF and MaxEnt models, and both RF and MaxEnt models had high predictive performance, with both mean AUC values of > 0.900 and all mean TSS values and Kappa values of > 0.800, and significant differences in the AUC (t = 19.862, P < 0.05), TSS (t = 10.140, P < 0.05) and Kappa values (t = 10.237, P < 0.05) between two models. The AUC, TSS and Kappa values of the ensemble model were 0.996, 0.954 and 0.920, respectively. Independent data verification showed that the AUC, TSS and Kappa values of the RF model and the ensemble model were all 1, which still showed high performance in unknown data modeling, and the MaxEnt model showed poor performance, with TSS and Kappa values of 0 for 24%(24/100) of the modeling results. The modeling results of 79 RF models, 38 MaxEnt models and their ensemble models with AUC values of > 0.950 and TSS values of > 0.850 were included in the evaluation of importance of environmental variables. The importance of annual sunshine duration (SSD) was 32.989%, 37.847% and 46.315% in the RF model, the MaxEnt model and their ensemble model, while the importance of annual mean relative humidity (RHU) was 30.947%, 15.921% and 28.121%, respectively. Important environment variables were concentrated in modeling results of the RF model, dispersed in modeling results of the MaxEnt model, and most concentrated in modeling results of the ensemble model. The potential distribution of O. hupensis snails after 2016 was predicted to be relatively concentrated in Yunnan Province by the RF model and relatively large by the MaxEnt model, and the distribution of O. hupensis snails predicted by the ensemble model was mostly the joint distribution of O. hupensis snails predicted by RF and MaxEnt models. Both RF and MaxEnt models are effective to predict the potential distribution of O. hupensis snails in Yunnan Province, which facilitates targeted O. hupensis snail control.

Predicting Post-Heart Transplant Composite Renal Outcome Risk in Adults: A Machine Learning Decision Tool

Assessing treatment switch among patients with multiple sclerosis: A machine learning approach

To investigate the value of machine learning(ML) in enhancing prostate cancer(PCa) diagnosis. Consecutive systematic prostate biopsies performed from Jan 2003-June 2017 were used as the training cohort, and prospective biopsies performed from July 2017-November 2019 were used as validation cohort. Men were included if PSA was 0.4-50 ng/mL, and information of digital rectal examination (DRE), Transrectal ultrasound(TRUS) prostate volume, TRUS abnormality were known. Clinically significant PCa(csPCa) was defined as Gleason 3 + 4 or above cancers. Area-under-curve (AUC) of receiver-operating characteristics (ROC) was compared between PSA, PSA density, European Randomized Study of Screening for Prostate Cancer (ERSPC) risk calculator (ERSPC-RC), and various ML techniques using PSA, DRE and TRUS information. ML techniques used included XGBoost, LightGBM, Catboost, Support vector machine (SVM), Logistic regression (LR), and Random Forest (RF), where cost sensitive learning was applied. Training and validation cohorts included 3881 and 778 consecutive men, respectively. RF model performed better than other ML techniques and PSA, PSA density and ERSPC-RC for prediction of PCa or csPCa in the validation cohort. In csPCa prediction, AUC of PSA, PSA density, ERSPC-RC and RF was 0.71, 0.80, 0.83 and 0.88 respectively. At 90-95% sensitivity for csPCa, RF model achieved a negative predictive value (NPV) of 97.5-98.0% and avoided 38.3-52.2% unnecessary biopsies. Decision curve analyses (DCA) showed RF model provided net clinical benefit over PSA, PSA density and ERSPC-RC. By using the same clinical parameters, ML techniques performed better than ERSPC-RC or PSA density in csPCa predictions, and could avoid up to 50% unnecessary biopsies.

Land-use regression model and machine learning algorithms in assessing outdoor noise exposure in Hong KongBackground/Aim: There are epidemiological studies in reporting the association between environmental noise exposure and various health outcomes, including cardiovascular diseases and cognitive decline. However, evidence in using land-use regression (LUR) model in predicting noise is limited and not conclusive, while machine learning algorithms are seemed to be a novel method in improving the explanative power of noise prediction. In this study, random forest (RF) model was adopted to estimate outdoor noise level in Hong Kong and compare the performance between LUR and RF. Methods: A total of 102 measurement sites with 27 environment-related variables were used to estimate the spatial environmental noise variation. Two separate measurements for A-weighted equivalent sound pressure levels over 24 hours (Leq,24h) were conducted at each site during the 2019-2020 period. Land use types, traffic information and geographic information systems were chosen to the related variables for building models. Ordinary linear regression was adopted to be the traditional land-use regression model. Results: The annual mean of 102 measurement sites of Leq,24h was 66.3 dB(A). The proportion of variation (R2) explained by the random forest model in full-frequency noise was higher than the traditional LUR (R2: RF: 0.76, LUR: 0.69). Random forest performed also better in lower root mean squared errors (RF: 2.91 dB(A), LUR: 3.22 dB(A)). Further results on the noise level in different time periods and frequencies will be described in the conference. Conclusion: Random forest model is suggested to be a better alternative in estimating noise exposure. Further study on the application of the model on epidemiological studies can be conducted.

BACKGROUNDFibrosis is a critical event in the progression of pediatric nonalcoholic fatty liver disease (NAFLD).AIMTo develop less invasive models based on machine learning (ML) to predict significant fibrosis in Chinese NAFLD children.METHODSIn this cross-sectional study, 222 and 101 NAFLD children with available liver biopsy data were included in the development of screening models for tertiary hospitals and community health centers, respectively. Predictive factors were selected using least absolute shrinkage and selection operator regression and stepwise logistic regression analyses. Logistic regression (LR) and other ML models were applied to construct the prediction models.RESULTSSimplified indicators of the ATS and BIU indices were constructed for tertiary hospitals and community health centers, respectively. When models based on the ATS and BIU parameter combinations were constructed, the random forest (RF) model demonstrated higher screening accuracy compared to the LR model (0.80 and 0.79 for the RF model and 0.72 and 0.77 for the LR model, respectively). Using cutoff values of 90% for sensitivity and 90% for specificity, the RF models could effectively identify and exclude NAFLD children with significant fibrosis in the internal validation set (with positive predictive values and negative prediction values exceeding 0.80), which could prevent liver biopsy in 60% and 71.4% of NAFLD children, respectively.CONCLUSIONThis study developed new models for predicting significant fibrosis in NAFLD children in tertiary hospitals and community health centers, which can serve as preliminary screening tools to detect the risk population in a timely manner.

BackgroundTranscutaneous transhepatic gallbladder drainage (PTGBD) has shown significant efficacy in the treatment of elderly patients with acute cholecystitis. The goal of this study is to develop a machine learning-based web calculator aimed at predicting the optimal timing for cholecystectomy (LC) after PTGBD in elderly patients with acute cholecystitis (AC) to achieve precise personalized medicine.MethodsA retrospective analysis of 979 elderly patients with acute cholecystitis admitted to Jinzhou Central Hospital and the First Affiliated Hospital of Jinzhou Medical University from 2013 to 2024 was performed, and a total of 680 patients were included in the model development. Patients were divided into delayed (347 cases, surgery > 6 weeks post-PTGBD) and non-delayed (333 cases) groups based on the interval between PTGBD and LC. Minimal Absolute Contraction and Selection Operator (LASSO) and logistic analysis were used to determine the predictors of postponement of LC in elderly patients with AC after PTGBD. Next, we used eight ML algorithms, namely Logistic Regression (LR), Decision Tree (DT), Random Forest (RF), Extreme gradient boosting (XGB), Support Vector Machine (SVM), Multilayer Perceptron (MLP), K-nearest Neighbor (KNN), Gaussian Naive Bayes (GNB), to train and develop ML models using a 10x cross-validation method. The performance of the model was evaluated by a variety of indicators, including the area under the receiver operating characteristic curve (ROC), calibration curve, decision curve, PR curve, and confusion matrix. In addition, model interpretation is performed through Shapley Additive Interpretation (SHAP) analysis to clarify the importance of each feature of the model and its basis for decision-making. Finally, we chose to use the best model to develop a web-based calculator that could be used to predict the likelihood of delaying LC after PTGBD in elderly AC patients.ResultsIn multivariate logistic regression analysis, age, sex, gallbladder wall thickness, time between onset and PTGBD, white blood cell count (WBC), C-reactive protein (CRP), and neutrophil-to-lymphocyte ratio (NLR) were identified as independent predictors of delayed LC in elderly patients with AC after PTGBD. In the training set, the area under the receiver operating characteristic curve (AUC) values for these models ranged from 0.808 to 0.914, with the random forest (RF) model showing the highest AUC value. Through the evaluation of decision curve analysis (DCA), precision-recall (PR) curve and calibration curve, the RF model showed superior clinical decision support and prediction performance compared with the other seven models. Finally, we used the RF model to build an online network calculator, which aims to accurately assist doctors in making more informed and accurate clinical decisions and promote the wide application of the model in clinical practice (https://zw17786325639.shinyapps.io/Postpone/).ConclusionsThis study developed and validated an RF model network calculator based on clinical indicator information to assess the likelihood of delaying LC after PTGBD in elderly patients with acute cholecystitis. This tool is expected to assist physicians in making more appropriate clinical decisions for patients.

Lung cancer remains the leading cause of cancer-related incidence and mortality worldwide. Its tendency for postoperative distant metastasis significantly compromises long-term prognosis and survival. Accurately predicting the metastatic potential in a timely manner is crucial for formulating optimal treatment strategies. This study aimed to comprehensively compare the predictive performance of nine machine learning (ML) models and to enhance interpretability through SHAP (Shapley Additive Explanations), with the goal of developing a practical and transparent risk stratification tool for postoperative lung cancer management. Clinical data from 3,120 patients with stage I-III lung cancer who underwent radical surgery were retrospectively collected and randomly divided into training and testing cohorts. A total of 52 clinical, pathological, imaging, and laboratory variables were analyzed. Nine ML models-including eXtreme Gradient Boosting (XGBoost), Random Forest (RF), Light Gradient Boosting Machine (LightGBM), Adaptive Boosting (AdaBoost), Decision Tree (DT), Gradient Boosting Decision Tree (GBDT), Gaussian Naive Bayes (GNB), Complement Naive Bayes (CNB), and Multilayer Perceptron classifier (MLP)-were developed and evaluated. Model performance was assessed using accuracy, precision, recall, F1 score, ROC-AUC, PR-AUC, calibration, and decision curve analysis (DCA). All models were evaluated using nested cross-validation (outer stratified 70/30 splits repeated 10 times; inner fivefold tuning), with decision thresholds prespecified in the inner loop and applied unchanged to held-out tests. Given the approximately 4:1 class imbalance, cost-sensitive learning was primarily adopted, and PR-AUC was reported in addition to ROC-AUC. Among the nine models, GBDT demonstrated the highest predictive performance, achieving an AUC of 0.810 (95% CI: 0.748-0.872), accuracy of 0.766, sensitivity of 0.698, and specificity of 0.786 in the test set. SHAP analysis revealed that adjuvant chemotherapy, adjuvant radiotherapy, pathological N stage, age, body mass index (BMI), and preoperative neutrophil count (Pre-ANC) were the most influential predictors of distant metastasis. The combination of model performance and interpretability supported the model's potential for integration into clinical workflows to assist in real-time decision-making. In this work, we carried out a systematic comparison of nine machine learning algorithms in a large postoperative cohort under a coherent and interpretable framework. By jointly considering discrimination, calibration, clinical benefit (via decision curve analysis), and SHAP-based explanations, we constructed a practical prognostic tool to guide personalized treatment strategies and follow-up care. This methodology offers a data-driven basis for precision management. Ultimately, our findings provide an internally validated reference framework that warrants external and multicenter validation prior to clinical deployment.

Background/Aims On average, there is a delay of 6.7 years between symptom onset and diagnosis of axial spondyloarthritis (axSpA). Since traditional approaches to improving early axSpA identification have had limited success, predictive automated analyses using patient records may help alleviate the burden on healthcare providers. We report results from a machine learning (ML) algorithm developed with UK electronic health records (EHRs) Clinical Practice Research Datalink (CPRD) data to estimate the probability or likelihood of a patient being diagnosed with axSpA based on prior clinical indicators and patient history. Methods Primary care UK EHR data - CPRD GOLD was used to identify patients with axSpA and healthy controls (HC). Patients aged ≥18 years with first diagnosis date of axSpA within the identification period (01-Jan-2005 to 31-Dec-2018) and fulfilling CPRD research acceptability criteria were included. Data pertaining to clinical presentation, consultation, referral, test, and therapy history were extracted for individual patients prior to diagnosis of axSpA. A total of 5,090 patients with axSpA satisfied the acceptability criteria. HC were randomly sampled to create a subset of one unique HC matched to each patient with axSpA, resulting in 5,089 HC. ML usable features derived from the total population (patients with axSpA and HC) numbered 820. After using a further exclusion criterion for the patients with axSpA and HC who had ≥1 of 820 usable features, the final dataset included 7,813 patients (3,902 with axSpA and 3,911 HC). This combined dataset was randomly split (67:33) into a train (n = 5237) and a test (n = 2576) dataset. A random forest (RF) model was trained on the train dataset. Cross-validation was performed for hyper-parameter tuning of the RF classifier. Once the model was trained, accuracy, precision, and F-1 scores were obtained with the test dataset. Results The RF-based algorithm resulted in a high level of accuracy (88.12%), with precision of 0.95 for patients with axSpA and 0.83 for HC. The RF algorithm identified 89 best clinical predictors (out of 820 used as inputs) that differentiated between patient and HC such as: total number of tests, total number of referrals, first age of consultation, first symptom age, and number of low back pain symptoms. The model sensitivity was 0.75 and positive predictive value was 80.88%. The model specificity was 0.96 and negative predictive value was 82.56%. Conclusion The ML algorithm demonstrated a high level of accuracy and precision in the identification of possible cases of axSpA, which may be useful in reducing the delay in diagnosis. Previous studies have successfully demonstrated automated cohort identification of axSpA in large datasets, with only a few using ML-based approaches for diagnosis from patient medical history. While our model supports previous work in axSpA, it needs further validation in routine clinical practice (exploration ongoing). Disclosure R. Sengupta: Honoraria; AbbVie, Biogen, Celgene, Lilly, MSD, Novartis, Roche, UCB. Grants/research support; AbbVie, Celgene, Novartis, UCB. S. Narasimham: Shareholder/stock ownership; Novartis. Other; Employee of Novartis. B.S. Mato: Shareholder/stock ownership; Novartis. Other; Employee of Novartis. M. Meglic: Other; Employee of Novartis. C. Perella: Other; Employee of Novartis. P. Pamies: Other; Employee of Novartis. P. Emery: Consultancies; AbbVie, Astra-Zeneca, BMS, Boehringer Ingelheim, Celltrion, Gilead, Janssen, MSD, Lilly, Novartis, Pfizer, Roche, Samsung, UCB. Grants/research support; AbbVie, BMS, Lilly, Novartis, Pfizer, Roche, Samsung.