External Validation Dataset Research Articles

#1017 External validation of a machine learning model for the prediction of bioimpedance-determined pre-dialysis overhydration

Abstract Background and Aims An accurate, automated determination of volume status in dialysis patients using a reproducible standard would be clinically useful. We have previously developed a machine learning model for the prediction of pre-dialysis overhydration (OH) index (litres) from the Fresenius Body Composition Monitor (BCM). External validation is a requisite step before the clinical implementation of a machine learning model. Method Data analyses were performed using Python. An XGBoost regressor model was internally developed and validated in a Fresenius dataset of 9,452 patients and 279,324 BCM target events from the Czech Republic and Italy between January 2016 and December 2022. An XGBoost regressor model with 60 input variables including the average BCM OH value in the preceding year, 2-week moving averages of dialysis session data and biomarkers/ medications data from the preceding 13 weeks demonstrated the best performance of the models tested. The external validation (EV) dataset consisted of 24 patients from a 7-week observational study in a single centre, in Beaumont, Ireland comprising 94 BCM sessions and 650 haemodialysis sessions. Individualised predictions were compared with BCM measures and clinical volume assessments performed at 2-week intervals. Mean squared error (MSE), Root mean square error (RMSE), and Bland-Altman plots were used for continuous outcomes. Precision, recall and F1 score were used to assess the algorithm's performance in predicting the following fluid categories: overhydration ≥1.1L, normohydration −1.1L to 1.1L, underhydration ≤-1.1L. Normohydration weight was calculated by subtracting the predicted OH index from the pre-dialysis weight. Results There was performance degradation between the internal validation (IV) and EV set. The RMSE for pre-dialysis OH was 1.04 kg in the IV set and 1.5 kg in the EV set. This could be attributed to differences in the distribution of the outcome variable (BCM OH validation dataset=1.97 ± 1.56 kg*, EV dataset=0.89 ± 1.7 kg), and the case-mix. The RMSE for normohydration weight was 1.86 kg in the EV set. Wide limits of agreement were noted on a Bland Altman plot comparing the average of BCM and predicted values plotted against the difference in values (mean bias - 0.54, 95% limits of agreement -3.44 kg to 2.36 kg). The overall accuracy of the model in predicting fluid classes was 60%. A histogram showed a close alignment in the distributions of predicted normohydration weights and BCM values. Conclusion The performance of the algorithms degraded between the internal validation and external validation datasets. However, the pre-dialysis algorithm showed an ability to discriminate fluid categories compared to nursing staff and the overall accuracy was acceptable. Additional training of the algorithms using fluid status classes would be expected to improve the precision of predictions. *Values less than the 1st percentile of – 2.17 L were excluded.

Improving the Generalizability and Performance of an Ultrasound Deep Learning Model Using Limited Multicenter Data for Lung Sliding Artifact Identification.

Deep learning (DL) models for medical image classification frequently struggle to generalize to data from outside institutions. Additional clinical data are also rarely collected to comprehensively assess and understand model performance amongst subgroups. Following the development of a single-center model to identify the lung sliding artifact on lung ultrasound (LUS), we pursued a validation strategy using external LUS data. As annotated LUS data are relatively scarce-compared to other medical imaging data-we adopted a novel technique to optimize the use of limited external data to improve model generalizability. Externally acquired LUS data from three tertiary care centers, totaling 641 clips from 238 patients, were used to assess the baseline generalizability of our lung sliding model. We then employed our novel Threshold-Aware Accumulative Fine-Tuning (TAAFT) method to fine-tune the baseline model and determine the minimum amount of data required to achieve predefined performance goals. A subgroup analysis was also performed and Grad-CAM++ explanations were examined. The final model was fine-tuned on one-third of the external dataset to achieve 0.917 sensitivity, 0.817 specificity, and 0.920 area under the receiver operator characteristic curve (AUC) on the external validation dataset, exceeding our predefined performance goals. Subgroup analyses identified LUS characteristics that most greatly challenged the model's performance. Grad-CAM++ saliency maps highlighted clinically relevant regions on M-mode images. We report a multicenter study that exploits limited available external data to improve the generalizability and performance of our lung sliding model while identifying poorly performing subgroups to inform future iterative improvements. This approach may contribute to efficiencies for DL researchers working with smaller quantities of external validation data.

Improving Targeted Mass Spectrometry Data Analysis with Nested Active Machine Learning

Targeted mass spectrometry (MS) holds promise for precise protein and protein‐representative peptide identification and quantification, enhancing disease diagnosis. However, its clinical application is hindered by complex data analysis and expert review requirements. It is hypothesized that machine learning (ML) models can automate data analysis to accelerate the clinical application of MS. The approach involves an ML‐driven pipeline that extracts statistical and morphological features from an MS target region and feeds these features into ML algorithms to generate and assess predictive models. The findings demonstrate ML prediction models exhibit superior performance when trained on extracted features versus raw spectra intensity data and that random forest models exhibit robust classification performance in both internal and external validation datasets. These models remain effective across varying training dataset sizes and positive sample rates and are enhanced by a nested active learning approach. This approach can thus revolutionize clinical MS applications.

Developing Parametric and Nonparametric Models for Model-Informed Precision Dosing: A Quality Improvement Effort in Vancomycin for Patients With Obesity.

Both parametric and nonparametric methods have been proposed to support model-informed precision dosing (MIPD). However, which approach leads to better models remains uncertain. Using open-source software, these 2 statistical approaches for model development were compared using the pharmacokinetics of vancomycin in a challenging subpopulation of class 3 obesity. Patients on vancomycin at the University of Vermont Medical Center from November 1, 2021, to February 14, 2023, were entered into the MIPD software. The inclusion criteria were body mass index (BMI) of at least 40 kg/m 2 and 1 or more vancomycin levels. A parametric model was created using nlmixr2/NONMEM, and a nonparametric model was created using Pmetrics. Then, a priori and a posteriori predictions were evaluated using the normalized root mean squared error (nRMSE) for precision and the mean percentage error (MPE) for bias. The parametric model was evaluated in a simulated MIPD context using an external validation dataset. In total, 83 patients were included in the model development, with a median age of 56.6 years (range: 24-89 years), and a median BMI of 46.3 kg/m 2 (range: 40-70.3 kg/m 2 ). Both parametric and nonparametric models were 2-compartmental, with creatinine clearance and fat-free mass as covariates to clearance and volume parameters, respectively. The a priori MPE and nRMSE for the parametric versus nonparametric models were -6.3% versus 2.69% and 27.2% versus 30.7%, respectively. The a posteriori MPE and RMSE were 0.16% and 0.84%, and 13.8% and 13.1%. The parametric model matched or outperformed previously published models on an external validation dataset (n = 576 patients). Minimal differences were found in the model structure and predictive error between the parametric and nonparametric approaches for modeling vancomycin class 3 obesity. However, the parametric model outperformed several other models, suggesting that institution-specific models may improve pharmacokinetics management.

Automatic segmentation of dura for quantitative analysis of lumbar stenosis: A deep learning study with 518 CT myelograms.

The diagnosis of lumbar spinal stenosis (LSS) can be challenging because radicular pain is not often present in the culprit-level localization. Accurate segmentation and quantitative analysis of the lumbar dura on radiographic images are key to the accurate differential diagnosis of LSS. The aim of this study is to develop an automatic dura-contouring tool for radiographic quantification on computed tomography myelogram (CTM) for patients with LSS. A total of 518 CTM cases with or without lumbar stenosis were included in this study. A deep learning (DL) segmentation algorithm 3-dimensional (3D) U-Net was deployed. A total of 210 labeled cases were used to develop the dura-contouring tool, with the ratio of the training, independent testing, and external validation datasets being 150:30:30. The Dice score (DCS) was the primary measure to evaluate the segmentation performance of the 3D U-Net, which was subsequently developed as the dura-contouring tool to segment another unlabeled 308 CTM cases with LSS. Automatic masks of 446 slices on the stenotic levels were then meticulously reviewed and revised by human experts, and the cross-sectional area (CSA) of the dura was compared. The mean DCS of the 3D U-Net were 0.905±0.080, 0.933±0.018, and 0.928±0.034 in the five-fold cross-validation, the independent testing, and the external validation datasets, respectively. The segmentation performance of the dura-contouring tool was also comparable to that of the second observer (the human expert). With the dura-contouring tool, only 59.0% (263/446) of the automatic masks of the stenotic slices needed to be revised. In the revised cases, there were no significant differences in the dura CSA between automatic masks and corresponding revised masks (p=0.652). Additionally, a strong correlation of dura CSA was found between the automatic masks and corresponding revised masks (r=0.805). A dura-contouring tool was developed that could automatically segment the dural sac on CTM, and it demonstrated high accuracy and generalization ability. Additionally, the dura-contouring tool has the potential to be applied in patients with LSS because it facilitates the quantification of the dural CSA on stenotic slices.

Development and Validation of an Explainable Deep Learning Model to Predict In-Hospital Mortality for Patients With Acute Myocardial Infarction: Algorithm Development and Validation Study.

Acute myocardial infarction (AMI) is one of the most severe cardiovascular diseases and is associated with a high risk of in-hospital mortality. However, the current deep learning models for in-hospital mortality prediction lack interpretability. This study aims to establish an explainable deep learning model to provide individualized in-hospital mortality prediction and risk factor assessment for patients with AMI. In this retrospective multicenter study, we used data for consecutive patients hospitalized with AMI from the Chongqing University Central Hospital between July 2016 and December 2022 and the Electronic Intensive Care Unit Collaborative Research Database. These patients were randomly divided into training (7668/10,955, 70%) and internal test (3287/10,955, 30%) data sets. In addition, data of patients with AMI from the Medical Information Mart for Intensive Care database were used for external validation. Deep learning models were used to predict in-hospital mortality in patients with AMI, and they were compared with linear and tree-based models. The Shapley Additive Explanations method was used to explain the model with the highest area under the receiver operating characteristic curve in both the internal test and external validation data sets to quantify and visualize the features that drive predictions. A total of 10,955 patients with AMI who were admitted to Chongqing University Central Hospital or included in the Electronic Intensive Care Unit Collaborative Research Database were randomly divided into a training data set of 7668 (70%) patients and an internal test data set of 3287 (30%) patients. A total of 9355 patients from the Medical Information Mart for Intensive Care database were included for independent external validation. In-hospital mortality occurred in 8.74% (670/7668), 8.73% (287/3287), and 9.12% (853/9355) of the patients in the training, internal test, and external validation cohorts, respectively. The Self-Attention and Intersample Attention Transformer model performed best in both the internal test data set and the external validation data set among the 9 prediction models, with the highest area under the receiver operating characteristic curve of 0.86 (95% CI 0.84-0.88) and 0.85 (95% CI 0.84-0.87), respectively. Older age, high heart rate, and low body temperature were the 3 most important predictors of increased mortality, according to the explanations of the Self-Attention and Intersample Attention Transformer model. The explainable deep learning model that we developed could provide estimates of mortality and visual contribution of the features to the prediction for a patient with AMI. The explanations suggested that older age, unstable vital signs, and metabolic disorders may increase the risk of mortality in patients with AMI.

Enhancing multiple myeloma staging: a novel cell death risk model approach

The prognostication of survival trajectories in multiple myeloma (MM) patients presents a substantial clinical challenge. Leveraging transcriptomic and clinical profiles from an expansive cohort of 2,088 MM patients, sourced from the Gene Expression Omnibus and The Cancer Genome Atlas repositories, we applied a sophisticated nested lasso regression technique to construct a prognostic model predicated on 28 gene pairings intrinsic to cell death pathways, thereby deriving a quantifiable risk stratification metric. Employing a threshold of 0.15, we dichotomized the MM samples into discrete high-risk and low-risk categories. Notably, the delineated high-risk cohort exhibited a statistically significant diminution in survival duration, a finding which consistently replicated across both training and external validation datasets. The prognostic acumen of our cell death signature was further corroborated by TIME ROC analyses, with the model demonstrating robust performance, evidenced by AUC metrics consistently surpassing the 0.6 benchmark across the evaluated arrays. Further analytical rigor was applied through multivariate COX regression analyses, which ratified the cell death risk model as an independent prognostic determinant. In an innovative stratagem, we amalgamated this risk stratification with the established International Staging System (ISS), culminating in the genesis of a novel, refined ISS categorization. This tripartite classification system was subjected to comparative analysis against extant prognostic models, whereupon it manifested superior predictive precision, as reflected by an elevated C-index. In summation, our endeavors have yielded a clinically viable gene pairing model predicated on cellular mortality, which, when synthesized with the ISS, engenders an augmented prognostic tool that exhibits pronounced predictive prowess in the context of multiple myeloma.

A novel artificial intelligence-based endoscopic ultrasonography diagnostic system for diagnosing the invasion depth of early gastric cancer

BackgroundWe developed an artificial intelligence (AI)-based endoscopic ultrasonography (EUS) system for diagnosing the invasion depth of early gastric cancer (EGC), and we evaluated the performance of this system.MethodsA total of 8280 EUS images from 559 EGC cases were collected from 11 institutions. Within this dataset, 3451 images (285 cases) from one institution were used as a development dataset. The AI model consisted of segmentation and classification steps, followed by the CycleGAN method to bridge differences in EUS images captured by different equipment. AI model performance was evaluated using an internal validation dataset collected from the same institution as the development dataset (1726 images, 135 cases). External validation was conducted using images collected from the other 10 institutions (3103 images, 139 cases).ResultsThe area under the curve (AUC) of the AI model in the internal validation dataset was 0.870 (95% CI: 0.796–0.944). Regarding diagnostic performance, the accuracy/sensitivity/specificity values of the AI model, experts (n = 6), and nonexperts (n = 8) were 82.2/63.4/90.4%, 81.9/66.3/88.7%, and 68.3/60.9/71.5%, respectively. The AUC of the AI model in the external validation dataset was 0.815 (95% CI: 0.743–0.886). The accuracy/sensitivity/specificity values of the AI model (74.1/73.1/75.0%) and the real-time diagnoses of experts (75.5/79.1/72.2%) in the external validation dataset were comparable.ConclusionsOur AI model demonstrated a diagnostic performance equivalent to that of experts.

MRI radiomics captures early treatment response in patient-derived organoid endometrial cancer mouse models.

Radiomics can capture microscale information in medical images beyond what is visible to the naked human eye. Using a clinically relevant mouse model for endometrial cancer, the objective of this study was to develop and validate a radiomic signature (RS) predicting response to standard chemotherapy. Mice orthotopically implanted with a patient-derived grade 3 endometrioid endometrial cancer organoid model (O-PDX) were allocated to chemotherapy (combined paclitaxel/carboplatin, n=11) or saline/control (n=13). During tumor progression, the mice underwent weekly T2-weighted (T2w) magnetic resonance imaging (MRI). Segmentation of primary tumor volume (vMRI) allowed extraction of radiomic features from whole-volume tumor masks. A radiomic model for predicting treatment response was derived employing least absolute shrinkage and selection operator (LASSO) statistics at endpoint images in the orthotopic O-PDX (RS_O), and subsequently applied on the earlier study timepoints (RS_O at baseline, and week 1-3). For external validation, the radiomic model was tested in a separate T2w-MRI dataset on segmented whole-volume subcutaneous tumors (RS_S) from the same O-PDX model, imaged at three timepoints (baseline, day 3 and day 10/endpoint) after start of chemotherapy (n=8 tumors) or saline/control (n=8 tumors). The RS_O yielded rapidly increasing area under the receiver operating characteristic (ROC) curves (AUCs) for predicting treatment response from baseline until endpoint; AUC=0.38 (baseline); 0.80 (week 1), 0.85 (week 2), 0.96 (week 3) and 1.0 (endpoint). In comparison, vMRI yielded AUCs of 0.37 (baseline); 0.69 (w1); 0.83 (week 2); 0.92 (week 3) and 0.97 (endpoint). When tested in the external validation dataset, RS_S yielded high accuracy for predicting treatment response at day10/endpoint (AUC=0.85) and tended to yield higher AUC than vMRI (AUC=0.78, p=0.18). Neither RS_S nor vMRI predicted response at day 3 in the external validation set (AUC=0.56 for both). We have developed and validated a radiomic signature that was able to capture chemotherapeutic treatment response both in an O-PDX and in a subcutaneous endometrial cancer mouse model. This study supports the promising role of preclinical imaging including radiomic tumor profiling to assess early treatment response in endometrial cancer models.

Biomarkers to improve functional outcome prediction after ischemic stroke: Results from the SICFAIL, STRAWINSKI, and PREDICT studies.

Acute ischemic stroke (AIS) outcome prognostication remains challenging despite available prognostic models. We investigated whether a biomarker panel improves the predictive performance of established prognostic scores. We investigated the improvement in discrimination, calibration, and overall performance by adding five biomarkers (procalcitonin, copeptin, cortisol, mid-regional pro-atrial natriuretic peptide (MR-proANP), and N-terminal pro-B-type natriuretic peptide (NT-proBNP)) to the Acute Stroke Registry and Analysis of Lausanne (ASTRAL) and age/NIHSS scores using data from two prospective cohort studies (SICFAIL, PREDICT) and one clinical trial (STRAWINSKI). Poor outcome was defined as mRS > 2 at 12 (SICFAIL, derivation dataset) or 3 months (PREDICT/STRAWINSKI, pooled external validation dataset). Among 412 SICFAIL participants (median age 70 years, quartiles 59-78; 63% male; median NIHSS score 3, quartiles 1-5), 29% had a poor outcome. Area under the curve of the ASTRAL and age/NIHSS were 0.76 (95% CI 0.71-0.81) and 0.77 (95% CI 0.73-0.82), respectively. Copeptin (0.79, 95% CI 0.74-0.84), NT-proBNP (0.80, 95% CI 0.76-0.84), and MR-proANP (0.79, 95% CI 0.75-0.84) significantly improved ASTRAL score's discrimination, calibration, and overall performance. Copeptin improved age/NIHSS model's discrimination, copeptin, MR-proANP, and NT-proBNP improved its calibration and overall performance. In the validation dataset (450 patients, median age 73 years, quartiles 66-81; 54% men; median NIHSS score 8, quartiles 3-14), copeptin was independently associated with various definitions of poor outcome and also mortality. Copeptin did not increase model's discrimination but it did improve calibration and overall model performance. Copeptin, NT-proBNP, and MR-proANP improved modest but consistently the predictive performance of established prognostic scores in patients with mild AIS. Copeptin was most consistently associated with poor outcome in patients with moderate to severe AIS, although its added prognostic value was less obvious.

One-stop detection of anterior cruciate ligament injuries on magnetic resonance imaging using deep learning with multicenter validation.

Anterior cruciate ligament (ACL) injuries are closely associated with knee osteoarthritis (OA). However, diagnosing ACL injuries based on knee magnetic resonance imaging (MRI) has been subjective and time-consuming for clinical doctors. Therefore, we aimed to devise a deep learning (DL) model leveraging MRI to enable a comprehensive and automated approach for the detection of ACL injuries. A retrospective study was performed extracting data from the Osteoarthritis Initiative (OAI). A total of 1,589 knees (comprising 1,443 intact, 90 with partial tears, and 56 with full tears) were enrolled to construct the classification model. This one-stop detection pipeline was developed using a tailored YOLOv5m architecture and a ResNet-18 convolutional neural network (CNN) to facilitate tasks based on sagittal 2-dimensional (2D) intermediate-weighted fast spin-echo sequence at 3.0T. To ensure the reliability and robustness of the classification system, it was subjected to external validation across 3 distinct datasets. The accuracy, sensitivity, specificity, and the mean average precision (mAP) were utilized as the evaluation metric for the model performance by employing a 5-fold cross-validation approach. The radiologist's interpretations were employed as the reference for conducting the evaluation. The localization model demonstrated an accuracy of 0.89 and a sensitivity of 0.93, achieving a mAP score of 0.96. The classification model demonstrated strong performance in detecting intact, partial tears, and full tears at the optimal threshold on the internal dataset, with sensitivities of 0.941, 0.833, and 0.929, specificities of 0.925, 0.947, and 0.991, and accuracies of 0.940, 0.941, and 0.989, respectively. In comparison, on a subset consisting of 171 randomly selected knees from the OAI, the radiologists demonstrated a sensitivity ranging between 0.660 and 1.000, specificity ranging between 0.691 and 1.000, and accuracy ranging between 0.689 and 1.000. On a subset consisting of 170 randomly selected knees from the Chinese dataset, the radiologists exhibited a sensitivity ranging between 0.711 and 0.948, specificity ranging between 0.768 and 0.977, and accuracy ranging between 0.683 and 0.917. After retraining, the model achieved sensitivities ranging between 0.630 and 0.961, specificities ranging between 0.860 and 0.961, and accuracies ranging between 0.832 and 0.951, respectively, on the external validation dataset. The proposed model utilizing knee MRI showcases robust performance in the domains of ACL localization and classification.

External validation of a nomogram for outcome prediction in management of medium-sized (1-2 cm) kidney stones.

Stone nomogram by Micali et al., able topredict treatment failure of shock-wave lithotripsy (SWL), retrograde intrarenal surgery (RIRS) and percutaneous nephrolithotomy (PNL) in the management of single 1-2 cm renal stones, was developed on 2605 patients and showed a high predictive accuracy, with an area under ROC curve of 0.793 at internal validation. The aim of the present study is to externally validate the model to assess whether it displayed a satisfactory predictive performance if applied to different populations. External validation was retrospectively performed on 3025 patients who underwent an active stone treatment from December 2010 to June 2021 in 26 centers from four countries (Italy, USA, Spain, Argentina). Collected variables included: age, gender, previous renal surgery, preoperative urine culture, hydronephrosis, stone side, site, density, skin-to-stone distance. Treatment failure was the defined outcome (residual fragments >4 mm at three months CT-scan). Model discrimination in external validation datasets showed an area under ROC curve of 0.66 (95% 0.59-0.68) with adequate calibration. The retrospective fashion of the study and the lack of generalizability of the tool towards populations from Asia, Africa or Oceania represent limitations of the current analysis. According to the current findings, Micali's nomogram can be used for treatment prediction after SWL, RIRS and PNL; however, a lower discrimination performance than the one at internal validation should be acknowledged, reflecting geographical, temporal and domain limitation of external validation studies. Further prospective evaluation is required to refine and improve the nomogram findings and to validate its clinical value.

Preoperative Risk Stratification for Gastric Cancer: The Establishment of Dual-Energy CT-Based Radiomics Using Prospective Datasets at Two Centers

ObjectiveTo evaluate the performance of dual-energy CT (DECT)-based radiomics models for identifying high-risk histopathologic phenotypes - serosal invasion (pT4a), lymph node metastasis (LNM), lymphovascular invasion (LVI) and perineural invasion (PNI) in gastric cancer. Material and MethodsThis prospective bi-center study recruited histologically confirmed gastric adenocarcinoma patients who underwent triple-phase enhanced DECT before gastrectomy between January 2021 and July 2023. Radiomics features were extracted from polychromatic/monochromatic (40keV, 100 keV)/iodine images at arterial/venous/delay phase, respectively. Predictive features were selected in the training dataset using logistic regression classifier, and trained models were applied to the external validation dataset. Performances of clinical models, conventional contrast enhanced CT (CECT) models and DECT models were evaluated using areas under the receiver operating characteristic curve (AUCs). ResultsIn total, 503 patients were recruited: 396 at training dataset (60.1 ± 10.8 years, 110 females, 286 males) and 107 at validation dataset (61.4 ± 9.5 years, 29 females, 78 males). DECT models dichotomizing pT4a, LNM, LVI, and PNI achieved AUCs of 0.891, 0.817, 0.834, and 0.889, respectively, in the validation dataset, similar with the CECT models. In the training dataset, compared with the CECT model, the DECT model provided increased performance for identifying pT4a, LNM, LVI (all P＜0.05), and similar performance for stratifying PNI (P = 0.104). The DECT models was associated with patient disease-free survival (all P ＜0.05). ConclusionDECT radiomics can stratify patients preoperatively according to high-risk histopathologic phenotypes for gastric cancer and are associated with patient disease-free survival in the training dataset.

Vision–language foundation model for echocardiogram interpretation

The development of robust artificial intelligence models for echocardiography has been limited by the availability of annotated clinical data. Here, to address this challenge and improve the performance of cardiac imaging models, we developed EchoCLIP, a vision–language foundation model for echocardiography, that learns the relationship between cardiac ultrasound images and the interpretations of expert cardiologists across a wide range of patients and indications for imaging. After training on 1,032,975 cardiac ultrasound videos and corresponding expert text, EchoCLIP performs well on a diverse range of benchmarks for cardiac image interpretation, despite not having been explicitly trained for individual interpretation tasks. EchoCLIP can assess cardiac function (mean absolute error of 7.1% when predicting left ventricular ejection fraction in an external validation dataset) and identify implanted intracardiac devices (area under the curve (AUC) of 0.84, 0.92 and 0.97 for pacemakers, percutaneous mitral valve repair and artificial aortic valves, respectively). We also developed a long-context variant (EchoCLIP-R) using a custom tokenizer based on common echocardiography concepts. EchoCLIP-R accurately identified unique patients across multiple videos (AUC of 0.86), identified clinical transitions such as heart transplants (AUC of 0.79) and cardiac surgery (AUC 0.77) and enabled robust image-to-text search (mean cross-modal retrieval rank in the top 1% of candidate text reports). These capabilities represent a substantial step toward understanding and applying foundation models in cardiovascular imaging for preliminary interpretation of echocardiographic findings.

Impact of magnetic resonance imaging-derived skeletal muscle index in locoregionally advanced nasopharyngeal carcinoma.

To evaluate the clinical implication of magnetic resonance imaging (MRI)-derived skeletal muscle index (SMI) in locoregionally advanced nasopharyngeal carcinoma (LANPC) patients undergoing induction chemotherapy (IC) followed by concurrent chemoradiotherapy (CCRT) and further to develop a nomogram for predicting survival prognosis. SMI was determined through baseline MRI at the third cervical level. The nomogram was based on a training cohort involving 409 LANPC patients. We validated the prognostic accuracy of this prognostic model in an internal validation cohort (n = 204) and an external independent cohort (n = 272). SMI was an independent risk factor for OS. A prognostic model comprising age, TNM stage and SMI for individual survival prediction was developed and graphically represented as a nomogram. The model showed favorable discrimination (C-index: 0.686), predictive accuracy [time dependent area under the curve (tAUC) at 5years: 0.70], and calibration, and was further validated in the internal and external validation datasets. A risk stratification derived from the model stratified these patients into three prognostic subgroups with significantly different survival. Low SMI accessed by MRI was significantly associated with poor overall survival in LANPC patients undergoing IC + CCRT. Moreover, we established and validated a novel nomogram involving age, TNM stage and SMI that could provide accurate prognostic stratification among this population.

Application of machine learning algorithms to identify people with low bone density.

Osteoporosis is becoming more common worldwide, imposing a substantial burden on individuals and society. The onset of osteoporosis is subtle, early detection is challenging, and population-wide screening is infeasible. Thus, there is a need to develop a method to identify those at high risk for osteoporosis. This study aimed to develop a machine learning algorithm to effectively identify people with low bone density, using readily available demographic and blood biochemical data. Using NHANES 2017-2020 data, participants over 50 years old with complete femoral neck BMD data were selected. This cohort was randomly divided into training (70%) and test (30%) sets. Lasso regression selected variables for inclusion in six machine learning models built on the training data: logistic regression (LR), support vector machine (SVM), gradient boosting machine (GBM), naive Bayes (NB), artificial neural network (ANN) and random forest (RF). NHANES data from the 2013-2014 cycle was used as an external validation set input into the models to verify their generalizability. Model discrimination was assessed via AUC, accuracy, sensitivity, specificity, precision and F1 score. Calibration curves evaluated goodness-of-fit. Decision curves determined clinical utility. The SHAP framework analyzed variable importance. A total of 3,545 participants were included in the internal validation set of this study, of whom 1870 had normal bone density and 1,675 had low bone density Lasso regression selected 19 variables. In the test set, AUC was 0.785 (LR), 0.780 (SVM), 0.775 (GBM), 0.729 (NB), 0.771 (ANN), and 0.768 (RF). The LR model has the best discrimination and a better calibration curve fit, the best clinical net benefit for the decision curve, and it also reflects good predictive power in the external validation dataset The top variables in the LR model were: age, BMI, gender, creatine phosphokinase, total cholesterol and alkaline phosphatase. The machine learning model demonstrated effective classification of low BMD using blood biomarkers. This could aid clinical decision making for osteoporosis prevention and management.

Machine-learning-based models to predict cardiovascular risk using oculomics and clinic variables in KNHANES

BackgroundRecent researches have found a strong correlation between the triglyceride-glucose (TyG) index or the atherogenic index of plasma (AIP) and cardiovascular disease (CVD) risk. However, there is a lack of research on non-invasive and rapid prediction of cardiovascular risk. We aimed to develop and validate a machine-learning model for predicting cardiovascular risk based on variables encompassing clinical questionnaires and oculomics.MethodsWe collected data from the Korean National Health and Nutrition Examination Survey (KNHANES). The training dataset (80% from the year 2008 to 2011 KNHANES) was used for machine learning model development, with internal validation using the remaining 20%. An external validation dataset from the year 2012 assessed the model’s predictive capacity for TyG-index or AIP in new cases. We included 32122 participants in the final dataset. Machine learning models used 25 algorithms were trained on oculomics measurements and clinical questionnaires to predict the range of TyG-index and AIP. The area under the receiver operating characteristic curve (AUC), accuracy, precision, recall, and F1 score were used to evaluate the performance of our machine learning models.ResultsBased on large-scale cohort studies, we determined TyG-index cut-off points at 8.0, 8.75 (upper one-third values), 8.93 (upper one-fourth values), and AIP cut-offs at 0.318, 0.34. Values surpassing these thresholds indicated elevated cardiovascular risk. The best-performing algorithm revealed TyG-index cut-offs at 8.0, 8.75, and 8.93 with internal validation AUCs of 0.812, 0.873, and 0.911, respectively. External validation AUCs were 0.809, 0.863, and 0.901. For AIP at 0.34, internal and external validation achieved similar AUCs of 0.849 and 0.842. Slightly lower performance was seen for the 0.318 cut-off, with AUCs of 0.844 and 0.836. Significant gender-based variations were noted for TyG-index at 8 (male AUC=0.832, female AUC=0.790) and 8.75 (male AUC=0.874, female AUC=0.862) and AIP at 0.318 (male AUC=0.853, female AUC=0.825) and 0.34 (male AUC=0.858, female AUC=0.831). Gender similarity in AUC (male AUC=0.907 versus female AUC=0.906) was observed only when the TyG-index cut-off point equals 8.93.ConclusionWe have established a simple and effective non-invasive machine learning model that has good clinical value for predicting cardiovascular risk in the general population.

Investigating potential biomarkers of acute pancreatitis in patients with a BMI>30 using Mendelian randomization and transcriptomic analysis

BackgroundAcute pancreatitis (AP) has become a significant global health concern, and a high body mass index (BMI) has been identified as a key risk factor exacerbating this condition. Within this context, lipid metabolism assumes a critical role. The complex relationship between elevated BMI and AP, mediated by lipid metabolism, markedly increases the risk of complications and mortality. This study aimed to accurately define the correlation between BMI and AP, incorporating a comprehensive analysis of the interactions between individuals with high BMI and AP.MethodsMendelian randomization (MR) analysis was first applied to determine the causal relationship between BMI and the risk of AP. Subsequently, three microarray datasets were obtained from the GEO database. This was followed by an analysis of differentially expressed genes and the application of weighted gene coexpression network analysis (WGCNA) to identify key modular genes associated with AP and elevated BMI. Functional enrichment analysis was then performed to shed light on disease pathogenesis. To identify the most informative genes, machine learning algorithms, including Random Forest (RF), Support Vector Machine-Recursive Feature Elimination (SVM-RFE), and Least Absolute Shrinkage and Selection Operator (LASSO), were employed. Subsequent analysis focused on the colocalization of the Quantitative Trait Loci (eQTL) data associated with the selected genes and Genome-Wide Association Studies (GWAS) data related to the disease. Preliminary verification of gene expression trends was conducted using external GEO datasets. Ultimately, the diagnostic potential of these genes was further confirmed through the development of an AP model in mice with a high BMI.ResultsA total of 21 intersecting genes related to BMI>30, AP, and lipid metabolism were identified from the datasets. These genes were primarily enriched in pathways related to cytosolic DNA sensing, cytokine‒cytokine receptor interactions, and various immune and inflammatory responses. Next, three machine learning techniques were utilized to identify HADH as the most prevalent diagnostic gene. Colocalization analysis revealed that HADH significantly influenced the risk factors associated with BMI and AP. Furthermore, the trend in HADH expression within the external validation dataset aligned with the trend in the experimental data, thus providing a preliminary validation of the experimental findings.The changes in its expression were further validated using external datasets and quantitative real-time polymerase chain reaction (qPCR).ConclusionThis study systematically identified HADH as a potential lipid metabolism-grounded biomarker for AP in patients with a BMI>30.

Identification of CT radiomic features robust to acquisition and segmentation variations for improved prediction of radiotherapy-treated lung cancer patient recurrence

The primary objective of the present study was to identify a subset of radiomic features extracted from primary tumor imaged by computed tomography of early-stage non-small cell lung cancer patients, which remain unaffected by variations in segmentation quality and in computed tomography image acquisition protocol. The robustness of these features to segmentation variations was assessed by analyzing the correlation of feature values extracted from lesion volumes delineated by two annotators. The robustness to variations in acquisition protocol was evaluated by examining the correlation of features extracted from high-dose and low-dose computed tomography scans, both of which were acquired for each patient as part of the stereotactic body radiotherapy planning process. Among 106 radiomic features considered, 21 were identified as robust. An analysis including univariate and multivariate assessments was subsequently conducted to estimate the predictive performance of these robust features on the outcome of early-stage non-small cell lung cancer patients treated with stereotactic body radiation therapy. The univariate predictive analysis revealed that robust features demonstrated superior predictive potential compared to non-robust features. The multivariate analysis indicated that linear regression models built with robust features displayed greater generalization capabilities by outperforming other models in predicting the outcomes of an external validation dataset.

A systematic evaluation of population pharmacokinetic models for polymyxin B in patients with liver and/or kidney dysfunction.

Polymyxin B (PMB) is considered a last-line treatment for multidrug-resistant (MDR) gram-negative bacterial infections. Model-informed precision dosing with population pharmacokinetics (PopPK) models could help to individualize PMB dosing regimens and improve therapy. However, the external prediction ability of the established PopPK models has not been fully elaborated. This study aimed to systemically evaluate eleven PMB PopPK models from ten published literature based on a new independent population, which was divided into four different populations, patients with liver dysfunction, kidney dysfunction, liver and kidney dysfunction, and normal liver and kidney function. The whole data set consisted of 146 patients with 391 PMB concentrations. The prediction- and simulation-based diagnostics and Bayesian forecasting were conducted to evaluate model predictability. In the overall evaluation process, none of the models exhibited satisfactory predictive ability in both prediction- and simulation-based diagnostic simultaneously. However, the evaluation of the models in the subgroup of patients with normal liver and kidney function revealed improved predictive performance compared to those with liver and/or kidney dysfunction. Bayesian forecasting demonstrated enhanced predictability with the incorporation of two to three prior observations. The external evaluation highlighted a lack of consistency between the prediction results of published models and the external validation dataset. Nonetheless, Bayesian forecasting holds promise in improving the predictive performance of the models, and feedback from therapeutic drug monitoring is crucial in optimizing individual dosing regimens.