Machine Learning-Based Prediction of Small Intracranial Aneurysm Rupture Status Using CTA-Derived Hemodynamics: A Multicenter Study.

A deep learning algorithm for detection of oral cavity squamous cell carcinoma from photographic images: A retrospective study.

IntroductionAsthma is one of the commonest chronic conditions in the world. Subtypes of asthma have been defined, typically from clinical datasets on small, well-characterised subpopulations of asthma patients. We sought to define asthma subtypes from large longitudinal primary care electronic health records (EHRs) using cluster analysis. MethodsIn this retrospective cohort study, we extracted asthma subpopulations from the Optimum Patient Care Research Database (OPCRD) to robustly train and test algorithms, and externally validated findings in the Secure Anonymised Information Linkage (SAIL) Databank. In both databases, we identified adults with an asthma diagnosis code recorded in the three years prior to an index date. Train and test datasets were selected from OPCRD using an index date of Jan 1, 2016. Two internal validation datasets were selected from OPCRD using index dates of Jan 1, 2017 and 2018. Three external validation datasets were selected from SAIL using index dates of Jan 1, 2016, 2017 and 2018. Each dataset comprised 50,000 randomly selected non-overlapping patients. Subtypes were defined by applying multiple correspondence analysis and k-means cluster analysis to the train dataset, and were validated in the internal and external validation datasets. ResultsWe defined six asthma subtypes with clear clinical interpretability: low inhaled corticosteroid (ICS) use and low healthcare utilisation (30% of patients); low-to-medium ICS use (36%); low-to-medium ICS use and comorbidities (12%); varied ICS use and comorbid chronic obstructive pulmonary disease (4%); high (10%) and very high ICS use (7%). The subtypes were replicated with high accuracy in internal (91–92%) and external (84–86%) datasets. ConclusionAsthma subtypes derived and validated in large independent EHR databases were primarily defined by level of ICS use, level of healthcare use, and presence of comorbidities. This has important clinical implications towards defining asthma subtypes, facilitating patient stratification, and developing more personalised monitoring and treatment strategies.

Lateral lymph node metastasis (LLNM) is very common in medullary thyroid carcinoma (MTC), but there is still controversy about how to manage cervical lateral lymph nodes, especially for clinically negative MTC. The aim of this study is to develop and validate a nomogram for predicting LLNM risk in MTC. A total of 234 patients from two hospitals were retrospectively enrolled in this study and divided into LLNM positive group and LLNM negative group based on the pathology. The correlation between LLNM and preoperative clinical and ultrasound variables were evaluated by univariable and multivariable logistic regression analysis. A nomogram was generated to predict the risk of the LLNM of MTC patients, validated by external dataset, and evaluated in terms of discrimination, calibration, and clinical usefulness. The training, internal, and external validation datasets included 152, 51, and 31 MTC patients, respectively. According to the multivariable logistic regression analysis, gender (male), relationship to thyroid capsule and serum calcitonin were independently associated with LLNM in the training dataset. The predictive nomogram model developed with the aforementioned variables showed favorable performance in estimating risk of LLNM, with the area under the ROC curve (AUC) of 0.826 in the training dataset, 0.816 in the internal validation dataset, and 0.846 in the external validation dataset. We developed and validated a model named MTC nomogram, utilizing available preoperative variables to predict the probability of LLNM in patients with MTC. This nomogram will be of great value for guiding the clinical diagnosis and treatment process of MTC patients.

The importance of body composition and sarcopenia is well-recognized in cancer patient outcomes and treatment tolerance, yet routine evaluations are rare due to their time-intensive nature. While CT scans provide accurate measurements, they depend on manual processes. We developed and validated a deep learning algorithm to automatically select and segment abdominal muscles [SM], visceral fat [VAT], and subcutaneous fat [SAT] on CT scans. A total of 352 CT scans were collected from two cancer centers. The detection of the third lumbar vertebra and three different body tissues (SM, VAT, and SAT) were annotated manually. The 5-fold cross-validation method was used to develop the algorithm and validate its performance on the training cohort. The results were validated on an external, independent group of CT scans. The algorithm for automatic L3 slice selection had a mean absolute error of 4 mm for the internal validation dataset and 5.5 mm for the external validation dataset. The median DICE similarity coefficient for body composition was 0.94 for SM, 0.93 for VAT, and 0.86 for SAT in the internal validation dataset, whereas it was 0.93 for SM, 0.93 for VAT, and 0.85 for SAT in the external validation dataset. There were high correlation scores with sarcopenia metrics in both internal and external validation datasets. Our deep learning algorithm facilitates routine research use and could be integrated into electronic patient records, enhancing care through better monitoring and the incorporation of targeted supportive measures like exercise and nutrition.

ObjectiveRobot-assisted partial nephrectomy (RAPN) is an established, minimally invasive technique to treat patients with renal masses. The incidence of acute kidney injury (AKI) after RAPN is high and is associated with poor prognosis. This study aims to develop and validate an interpretable machine-learning model based on clinical features for individualized risk assessment of RAPN-AKI.MethodsWe retrospectively reviewed 325 patients undergoing RAPN at the Third Medical Center of PLA General Hospital (May 2022–Oct 2023) as the training dataset, and 146 from the Fifth Medical Center of PLA General Hospital (Nov 2023–Dec 2024) for external validation. Models were constructed using Boruta-selected features and eight machine learning algorithms. Performance was assessed by the area under the receiver operating characteristic curve (AUC), F1-score, accuracy, precision, calibration, and decision curve analysis (DCA). Shapley additive explanations (SHAP) interpreted feature contributions.ResultsThe incidence of AKI in internal training and external validation datasets was 24.6% and 26%, respectively. The Boruta algorithm identified duration of renal artery blockade, preoperative serum creatinine (Scr), gender, body mass index (BMI), and age as important features. Among the eight machine learning models, the Gradient Boosting Machine (GBM) model demonstrated the best and most stable predictive outcomes in the internal training dataset (AUC = 0.889) and external validation dataset (AUC = 0.779). Both the calibration curve and DCA indicated better calibration and greater net benefit. SHAP analysis revealed the contribution of important features in the following order: duration of renal artery blockade, Scr, BMI, age, and gender. Dependency plots showed that duration of renal artery blockade > 22 min, Scr > 80 µmol/L, BMI > 25 kg/m², age > 60 years, and male were significantly associated with an increased risk of AKI.ConclusionThe GBM model exhibited strong predictive performance in both internal training dataset and external validation dataset and has the potential to assist clinicians to identify the high-risk patients early, enabling timely interventions that may reduce the incidence of RAPN-AKI and improving clinical outcomes. While, the interpretable machine learning model is currently applicable only to patients with low-risk or normal preoperative renal function.

INTRODUCTION:Conventional white light imaging (WLI) endoscopy is the most common screening technique used for detecting early esophageal squamous cell carcinoma (ESCC). Nevertheless, it is difficult to detect and delineate margins of early ESCC using WLI endoscopy. This study aimed to develop an artificial intelligence (AI) model to detect and delineate margins of early ESCC under WLI endoscopy.METHODS:A total of 13,083 WLI images from 1,239 patients were used to train and test the AI model. To evaluate the detection performance of the model, 1,479 images and 563 images were used as internal and external validation data sets, respectively. For assessing the delineation performance of the model, 1,114 images and 211 images were used as internal and external validation data sets, respectively. In addition, 216 images were used to compare the delineation performance between the model and endoscopists.RESULTS:The model showed an accuracy of 85.7% and 84.5% in detecting lesions in internal and external validation, respectively. For delineating margins, the model achieved an accuracy of 93.4% and 95.7% in the internal and external validation, respectively, under an overlap ratio of 0.60. The accuracy of the model, senior endoscopists, and expert endoscopists in delineating margins were 98.1%, 78.6%, and 95.3%, respectively. The proposed model achieved similar delineating performance compared with that of expert endoscopists but superior to senior endoscopists.DISCUSSION:We successfully developed an AI model, which can be used to accurately detect early ESCC and delineate the margins of the lesions under WLI endoscopy.

Early diagnosis of postmenopausal osteoporosis provides an opportunity to detect and prevent fractures. This study uses machine learning (ML) techniques to enhance the predictive ability for low bone mass (LBM) risk. A retrospective cross-sectional study was performed, including 3,738 menopausal women from a hospital (the internal validation dataset) and 1,008 menopausal women from the community (the external validation dataset) between December 2014 and February 2022. The least absolute shrinkage and selection operation (LASSO) and elastic net methods are employed to screen the variables. ML algorithms and logistic regression are applied using clinical risk factors to develop a prediction model, and its effectiveness is subsequently evaluated. The optimal model is selected, and the concordance statistic is established for discrimination, comprising 11 variables. In predicting LBM, the model achieves an AUC of 0.918 in the internal validation dataset and 0.910 in the external validation dataset, with the XGboost model particularly noteworthy. This prediction model assists older women at elevated risk of osteoporosis, guiding decision-making for primary care providers to identify those needing preventive treatment.

The aim of this study was to develop and validate a deep learning-based algorithm (DLA) for automatic detection and grading of motion-related artifacts on arterial phase liver magnetic resonance imaging (MRI). Multistep DLA for detection and grading of motion-related artifacts, based on the modified ResNet-101 and U-net, were trained using 336 arterial phase images of gadoxetic acid-enhanced liver MRI examinations obtained in 2017 (training dataset; mean age, 68.6 years [range, 18-95]; 254 men). Motion-related artifacts were evaluated in 4 different MRI slices using a 3-tier grading system. In the validation dataset, 313 images from the same institution obtained in 2018 (internal validation dataset; mean age, 67.2 years [range, 21-87]; 228 men) and 329 from 3 different institutions (external validation dataset; mean age, 64.0 years [range, 23-90]; 214 men) were included, and the per-slice and per-examination performances for the detection of motion-related artifacts were evaluated. The per-slice sensitivity and specificity of the DLA for detecting grade 3 motion-related artifacts were 91.5% (97/106) and 96.8% (1134/1172) in the internal validation dataset and 93.3% (265/284) and 91.6% (948/1035) in the external validation dataset. The per-examination sensitivity and specificity were 92.0% (23/25) and 99.7% (287/288) in the internal validation dataset and 90.0% (72/80) and 96.0% (239/249) in the external validation dataset, respectively. The processing time of the DLA for automatic grading of motion-related artifacts was from 4.11 to 4.22 seconds per MRI examination. The DLA enabled automatic and instant detection and grading of motion-related artifacts on arterial phase gadoxetic acid-enhanced liver MRI.

Non-invasive biopsy diagnosis of diabetic kidney disease via deep learning applied to retinal images: a population-based study.

In the Anatomical Therapeutic Chemical (ATC) classification system, therapeutic drugs are divided into 14 main classes according to the organ or system on which they act and their chemical, pharmacological and therapeutic properties. This system, recommended by the World Health Organization (WHO), provides a global standard for classifying medical substances and serves as a tool for international drug utilization research to improve quality of drug use. In view of this, it is necessary to develop effective computational prediction methods to identify the ATC-class of a given drug, which thereby could facilitate further analysis of this system. In this study, we initiated an attempt to develop a prediction method and to gain insights from it by utilizing ontology information of drug compounds. Since only about one-fourth of drugs in the ATC classification system have ontology information, a hybrid prediction method combining the ontology information, chemical interaction information and chemical structure information of drug compounds was proposed for the prediction of drug ATC-classes. As a result, by using the Jackknife test, the 1st prediction accuracies for identifying the 14 main ATC-classes in the training dataset, the internal validation dataset and the external validation dataset were 75.90%, 75.70% and 66.36%, respectively. Analysis of some samples with false-positive predictions in the internal and external validation datasets indicated that some of them may even have a relationship with the false-positive predicted ATC-class, suggesting novel uses of these drugs. It was conceivable that the proposed method could be used as an efficient tool to identify ATC-classes of novel drugs or to discover novel uses of known drugs.

Deep learning (DL) algorithms could improve the classification of ovarian tumours assessed with ultrasound (US) images and clinical information. This study aimed to develop a DL model for diagnosing ovarian tumours based on US images and clinical information, and to compare the performance of the DL model with radiologist assessment. This study retrospectively collected data from four hospitals involving women who underwent US examinations for ovarian tumours. Additionally, data were prospectively randomly collected from two other hospitals. Histopathological analysis served as the reference standard. The retrospective dataset was divided into training, test and internal validation sets. The prospective dataset was used as an external validation set. A DL model was developed. The performance of the DL model was compared with the eleven radiologists’ assessments in the internal and external validation sets. The performance of radiologists' assessments with or without DL model were also compared. A total of 1518 women (6432 images) with ovarian tumours were collected, including 364 women in the internal validation dataset and 340 women in the external validation dataset. In the external validation dataset, the accuracy for the DL model's classification of benign, borderline, and malignant ovarian tumours were 85.5%, which were comparable to the expert assessments (85.3%, p > 0.05), but significantly higher than the average level of eleven different-experience radiologists (72.5%, p < 0.05). The diagnostic performance of junior and mid-level experienced radiologists were significantly improved after applying the DL model. The DL model developed using US images and clinical information can effectively classify benign, malignant, and borderline ovarian tumours with diagnostic performance comparable to the expert assessment and improving junior and mid-level experienced radiologists.

BackgroundAccurate evaluation of glomerular filtration rates (GFRs) is of critical importance in clinical practice. A previous study showed that models based on artificial neural networks (ANNs) could achieve a better performance than traditional equations. However, large-sample cross-sectional surveys have not resolved questions about ANN performance.MethodsA total of 1,180 patients that had chronic kidney disease (CKD) were enrolled in the development data set, the internal validation data set and the external validation data set. Additional 222 patients that were admitted to two independent institutions were externally validated. Several ANNs were constructed and finally a Back Propagation network optimized by a genetic algorithm (GABP network) was chosen as a superior model, which included six input variables; i.e., serum creatinine, serum urea nitrogen, age, height, weight and gender, and estimated GFR as the one output variable. Performance was then compared with the Cockcroft-Gault equation, the MDRD equations and the CKD-EPI equation.ResultsIn the external validation data set, Bland-Altman analysis demonstrated that the precision of the six-variable GABP network was the highest among all of the estimation models; i.e., 46.7 ml/min/1.73 m2 vs. a range from 71.3 to 101.7 ml/min/1.73 m2, allowing improvement in accuracy (15% accuracy, 49.0%; 30% accuracy, 75.1%; 50% accuracy, 90.5% [P<0.001 for all]) and CKD stage classification (misclassification rate of CKD stage, 32.4% vs. a range from 47.3% to 53.3% [P<0.001 for all]). Furthermore, in the additional external validation data set, precision and accuracy were improved by the six-variable GABP network.ConclusionsA new ANN model (the six-variable GABP network) for CKD patients was developed that could provide a simple, more accurate and reliable means for the estimation of GFR and stage of CKD than traditional equations. Further validations are needed to assess the ability of the ANN model in diverse populations.

We aimed to develop an effective model to identify the risk of concurrent bacterial infections in older patients with acute-on-chronic liver disease (AoCLD). Data from 809 individuals aged 60-80 sourced from the CATCH-LIFE cohort were analyzed. Participants were randomly assigned to training and internal validation groups at a ratio of 7:3. An independent cohort of 336 older inpatients with AoCLD from Xiangya Hospital, Central South University was used to conduct an external validation of the model. Independent risk factors were identified using LASSO and logistic regression analysis in the training cohort and were subsequently used to develop a user-friendly model. Model performance was evaluated using area under the curve (AUC), calibration plots, and decision curve analysis in the internal and external validation cohorts. Two different cutoff values were determined to stratify infection risk in older patients with AoCLD. The infection rate among older patients with AoCLD was 30.28%. Pulmonary infections were predominant, accounting for 93% of all infections. Gram-negative bacteria were the most frequently isolated pathogens, representing 64% of cases in this population. The novel model developed to identify bacterial infections included three variables: cirrhosis, absolute neutrophil count, and C-reactive protein (CRP) level. The AUC for the training, internal, and external validation datasets demonstrated high accuracy in identifying bacterial infections (AUC of the training dataset = 0.805, AUC of the internal validation dataset = 0.848, and AUC of the external validation dataset = 0.838). The model significantly outperformed neutrophil count, CRP level, and procalcitonin level alone in detecting bacterial infections among older patients with AoCLD. To facilitate clinical decision-making, we defined two cutoff values of prediction probability: a low cutoff of 32.2% to rule out bacterial infections and a high cutoff of 47.9% to confidently confirm bacterial infections. Our model aids in the early and precise diagnosis of bacterial infections in older patients with AoCLD, thereby facilitating prompt interventions to prevent adverse outcomes.

SummarySustained engagement in HIV care and adherence to antiretroviral therapy (ART) are essential for achieving the UNAIDS “95-95-95” targets. Despite increased ART access, disengagement from care remains a significant issue, particularly in sub-Saharan Africa. Traditional machine learning (ML) models have shown moderate success in predicting care disengagement, which would enable early intervention. We develop an enhanced large language model (LLM) fine-tuned with electronic medical records (EMRs) to predict people at risk of disengaging from HIV care in Tanzania and to provide interpretative insights into modifiable risk factors.MethodsWe developed a novel AI model by enhancing a pre-trained LLM (LLaMA 3.1, an open-source pre-trained LLM released by Meta) using routinely collected EMRs from Tanzania’s National HIV Care and Treatment Program from January 1, 2018, to June 30, 2023 (4,809,765 records for 261,192 people) to identify people at risk of disengaging from HIV care or developing adverse outcomes. Outcomes included risk of ART non-adherence, non-suppressed viral load, and loss to follow-up. Models were evaluated internally (Kagera region) and externally (Geita region), with performance compared against state-of-art ML models and zero-shot LLMs. Additionally, a team of HIV physicians in Tanzania assessed the LLM’s predictions along with LLM provided justifications for a subset of patient records to evaluate their clinical relevance and reasoning.FindingsThe enhanced LLMs consistently outperformed the supervised ML model and zero-shot LLMs across all outcomes in both internal and external validation datasets. When focusing on the 25% of PLHIV predicted as most likely to lost-to-follow-up (LTFU), the model correctly identified 78% (2,515 of 3,224) of people living with HIV (PLHIV) genuinely at risk in internal validation and 73% (7,105 of 9,733) in external validation. Attention score analysis indicated that the enhanced LLM focused on keywords such as gaps in follow-up care and ART adherence. The human expert evaluation showed alignment between clinician assessments and the LLM’s predictions in 65% of cases, with experts finding the model’s justifications reasonable and clinically relevant in 92.3% of aligned cases.InterpretationIf implemented in HIV clinics, this LLM-based AI model could help allocate resources efficiently and deliver targeted interventions, improving retention in care and advancing the UNAIDS “95-95-95” targets. By functioning like a clinician–analyzing patient summaries, predicting risks, and offering reasoning–the enhanced LLM could be integrated into clinical workflows to complement human expertise, facilitating timely interventions and informed decision-making. If implemented widely, this human-AI collaboration has the potential to improve health outcomes for people living with HIV and reduce onward transmission.FundingThe study was supported by a grant from the US National Institutes of Health (NIH): NIH NIMH 1R01MH125746.

The diagnosis of thyroid-associated ophthalmopathy (TAO) usually requires a comprehensive examination, including clinical symptoms, radiological examinations, and blood tests. Therefore, cost-effective and noninvasive methods for the detection of TAO are needed. This study aimed to establish a deep learning-based system to detect TAO based on retinal photographs. The multicenter observational study included retinal photographs taken from TAO patients and normal participants in 2 hospitals in China. Forty-five-degree retinal photographs, centered on the midpoint between the optic disc and the macula, were captured by trained ophthalmologists. The authors first trained a convolutional neural network model to identify TAO using data collected from one hospital. After internal validation, the model was further evaluated in another hospital as an external validation data set. The study included 1182 retinal photographs of 708 participants for model development, and 365 retinal photographs (189 participants) were obtained as the external validation data set. In the internal validation, the area under the receiver operator curve was 0.900 (95% CI: 0.889-0.910) and the accuracy was 0.860 (95% CI: 0.849-0.869). In the external data set, the model reached an area under the curve of 0.747 (95% CI: 0.728-0.763) and achieved an accuracy of 0.709 (95% CI: 0.690-0.724). Deep learning-based systems may be promising for identifying TAO in normal subjects using retinal fundus photographs. It may serve as a cost-effective and noninvasive method to detect TAO in the future.