Creating realistic anterior segment optical coherence tomography images using generative adversarial networks.

AIMS: To develop a generative adversarial network (GAN) capable of generating realistic high-resolution anterior segment optical coherence tomography (AS-OCT) images. METHODS: This study included 142 628 AS-OCT B-scans from the American University of Beirut Medical Center. The Style and WAvelet based GAN architecture was trained to generate realistic AS-OCT images and was evaluated through the Fréchet Inception Distance (FID) Score and a blinded assessment by three refractive surgeons who were asked to distinguish between real and generated images. To assess the suitability of the generated images for machine learning tasks, a convolutional neural network (CNN) was trained using a dataset of real and generated images over a classification task. The generated AS-OCT images were then upsampled using an enhanced super-resolution GAN (ESRGAN) to achieve high resolution. RESULTS: The generated images exhibited visual and quantitative similarity to real AS-OCT images. Quantitative similarity assessed using FID scored an average of 6.32. Surgeons scored 51.7% in identifying real versus generated images which was not significantly better than chance (p value >0.3). The CNN accuracy improved from 78% to 100% when synthetic images were added to the dataset. The ESRGAN upsampled images were objectively more realistic and accurate compared with traditional upsampling techniques by scoring a lower Learned Perceptual Image Patch Similarity of 0.0905 compared with 0.4244 of bicubic interpolation. CONCLUSIONS: This study successfully developed and leveraged GANs capable of generating high-definition synthetic AS-OCT images that are realistic and suitable for machine learning and image analysis tasks.

Almanac - Retrieval-Augmented Language Models for Clinical Medicine.

BACKGROUND: Large language models (LLMs) have recently shown impressive zero-shot capabilities, whereby they can use auxiliary data, without the availability of task-specific training examples, to complete a variety of natural language tasks, such as summarization, dialogue generation, and question answering. However, despite many promising applications of LLMs in clinical medicine, adoption of these models has been limited by their tendency to generate incorrect and sometimes even harmful statements. METHODS: We tasked a panel of eight board-certified clinicians and two health care practitioners with evaluating Almanac, an LLM framework augmented with retrieval capabilities from curated medical resources for medical guideline and treatment recommendations. The panel compared responses from Almanac and standard LLMs (ChatGPT-4, Bing, and Bard) versus a novel data set of 314 clinical questions spanning nine medical specialties. RESULTS: Almanac showed a significant improvement in performance compared with the standard LLMs across axes of factuality, completeness, user preference, and adversarial safety. CONCLUSIONS: Our results show the potential for LLMs with access to domain-specific corpora to be effective in clinical decision-making. The findings also underscore the importance of carefully testing LLMs before deployment to mitigate their shortcomings. (Funded by the National Institutes of Health, National Heart, Lung, and Blood Institute.).

Machine Learning Multicenter Risk Model to Predict Right Ventricular Failure After Mechanical Circulatory Support: The STOP-RVF Score.

Importance: The existing models predicting right ventricular failure (RVF) after durable left ventricular assist device (LVAD) support might be limited, partly due to lack of external validation, marginal predictive power, and absence of intraoperative characteristics. Objective: To derive and validate a risk model to predict RVF after LVAD implantation. Design, Setting, and Participants: This was a hybrid prospective-retrospective multicenter cohort study conducted from April 2008 to July 2019 of patients with advanced heart failure (HF) requiring continuous-flow LVAD. The derivation cohort included patients enrolled at 5 institutions. The external validation cohort included patients enrolled at a sixth institution within the same period. Study data were analyzed October 2022 to August 2023. Exposures: Study participants underwent chronic continuous-flow LVAD support. Main Outcome and Measures: The primary outcome was RVF incidence, defined as the need for RV assist device or intravenous inotropes for greater than 14 days. Bootstrap imputation and adaptive least absolute shrinkage and selection operator variable selection techniques were used to derive a predictive model. An RVF risk calculator (STOP-RVF) was then developed and subsequently externally validated, which can provide personalized quantification of the risk for LVAD candidates. Its predictive accuracy was compared with previously published RVF scores. Results: The derivation cohort included 798 patients (mean [SE] age, 56.1 [13.2] years; 668 male [83.7%]). The external validation cohort included 327 patients. RVF developed in 193 of 798 patients (24.2%) in the derivation cohort and 107 of 327 patients (32.7%) in the validation cohort. Preimplant variables associated with postoperative RVF included nonischemic cardiomyopathy, intra-aortic balloon pump, microaxial percutaneous left ventricular assist device/venoarterial extracorporeal membrane oxygenation, LVAD configuration, Interagency Registry for Mechanically Assisted Circulatory Support profiles 1 to 2, right atrial/pulmonary capillary wedge pressure ratio, use of angiotensin-converting enzyme inhibitors, platelet count, and serum sodium, albumin, and creatinine levels. Inclusion of intraoperative characteristics did not improve model performance. The calculator achieved a C statistic of 0.75 (95% CI, 0.71-0.79) in the derivation cohort and 0.73 (95% CI, 0.67-0.80) in the validation cohort. Cumulative survival was higher in patients composing the low-risk group (estimated <20% RVF risk) compared with those in the higher-risk groups. The STOP-RVF risk calculator exhibited a significantly better performance than commonly used risk scores proposed by Kormos et al (C statistic, 0.58; 95% CI, 0.53-0.63) and Drakos et al (C statistic, 0.62; 95% CI, 0.57-0.67). Conclusions and Relevance: Implementing routine clinical data, this multicenter cohort study derived and validated the STOP-RVF calculator as a personalized risk assessment tool for the prediction of RVF and RVF-associated all-cause mortality.

Deep Learning-Based Estimation of Implantable Collamer Lens Vault Using Optical Coherence Tomography.

PURPOSE: To develop and validate a deep learning neural network for automated measurement of implantable collamer lens (ICL) vault using anterior segment optical coherence tomography (AS-OCT). DESIGN: Cross-sectional retrospective study. METHODS: A total of 2647 AS-OCT scans were used from 139 eyes of 82 subjects who underwent ICL surgery in 3 different centers. Using transfer learning, a deep learning network was trained and validated for estimating the ICL vault on OCT. A trained operator separately reviewed all OCT scans and measured the central vault using a built-in caliper tool. The model was then separately tested on 191 scans. A Bland-Altman plot was constructed and the mean absolute percentage error (MAPE), mean absolute error (MAE), root mean squared error (RMSE), Pearson correlation coefficient (r), and determination coefficient (R2) were calculated to evaluate the strength and validity of the model. RESULTS: On the test set, the model achieved a MAPE of 3.42%, an MAE of 15.82 µm, a RMSE of 18.85 µm, a Pearson correlation coefficient r of +0.98 (P < .00001), and a coefficient of determination R2 of +0.96. There was no significant difference between the vaults of the test set labeled by the technician vs those estimated by the model: 478 ± 95 µm vs 475 ± 97 µm, respectively, P = .064). CONCLUSIONS: Using transfer learning, our deep learning neural network was able to accurately compute the ICL vault from AS-OCT scans, overcoming the limitations of an imbalanced data set and limited training data. Such an algorithm can assist the postoperative assessment in ICL surgery.

Almanac: Retrieval-Augmented Language Models for Clinical Medicine.

Large-language models have recently demonstrated impressive zero-shot capabilities in a variety of natural language tasks such as summarization, dialogue generation, and question-answering. Despite many promising applications in clinical medicine, adoption of these models in real-world settings has been largely limited by their tendency to generate incorrect and sometimes even toxic statements. In this study, we develop Almanac, a large language model framework augmented with retrieval capabilities for medical guideline and treatment recommendations. Performance on a novel dataset of clinical scenarios (n= 130) evaluated by a panel of 5 board-certified and resident physicians demonstrates significant increases in factuality (mean of 18% at p-value < 0.05) across all specialties, with improvements in completeness and safety. Our results demonstrate the potential for large language models to be effective tools in the clinical decision-making process, while also emphasizing the importance of careful testing and deployment to mitigate their shortcomings.

Predicting Dose-Range Chemical Toxicity using Novel Hybrid Deep Machine-Learning Method.

Humans are exposed to thousands of chemicals, including environmental chemicals. Unfortunately, little is known about their potential toxicity, as determining the toxicity remains challenging due to the substantial resources required to assess a chemical in vivo. Here, we present a novel hybrid neural network (HNN) deep learning method, called HNN-Tox, to predict chemical toxicity at different doses. To develop a hybrid HNN-Tox method, we combined two neural network frameworks, the Convolutional Neural Network (CNN) and the multilayer perceptron (MLP)-type feed-forward neural network (FFNN). Combining the CNN and FCNN in the field of environmental chemical toxicity prediction is a novel approach. We developed several binary and multiclass classification models to assess dose-range chemical toxicity that is trained based on thousands of chemicals with known toxicity. The performance of the HNN-Tox was compared with other machine-learning methods, including Random Forest (RF), Bootstrap Aggregation (Bagging), and Adaptive Boosting (AdaBoost). We also analyzed the model performance dependency on varying features, descriptors, dataset size, route of exposure, and toxic dose. The HNN-Tox model, trained on 59,373 chemicals annotated with known LD50 and routes of exposure, maintained its predictive ability with an accuracy of 84.9% and 84.1%, even after reducing the descriptor size from 318 to 51, and the area under the ROC curve (AUC) was 0.89 and 0.88, respectively. Further, we validated the HNN-Tox with several external toxic chemical datasets on a large scale. The HNN-Tox performed optimally or better than the other machine-learning methods for diverse chemicals. This study is the first to report a large-scale prediction of dose-range chemical toxicity with varying features. The HNN-Tox has broad applicability in predicting toxicity for diverse chemicals and could serve as an alternative methodology approach to animal-based toxicity assessment.

Treatment-induced cerebral sinus venous thrombosis in childhood acute lymphoblastic malignancies: New risk factors to consider.

BACKGROUND: Cerebral sinus venous thrombosis (CSVT) is one of the many side effects encountered during acute lymphoblastic leukemia (ALL) therapy. Due to the rarity of cases, lack of data, and consensus management, no recommendations exist to target the population at risk. METHODS: This is a retrospective chart review of 229 consecutive patients diagnosed with ALL with an age range of 1-21 years, treated at the Children's Cancer Center of Lebanon between October 2007 and February 2018. RESULTS: The incidence of CSVT was 10.5%. Using univariate analysis, increased risk of CSVT was observed with male gender, age >10 years, T-cell immunophenotype, intermediate/high-risk disease, maximum triglyceride (TG) level of >615 mg/dl, presence of mediastinal mass, and larger body surface area (BSA). With multivariate analysis, the only statistically significant risk factors were maximum TG level, BSA, presence of mediastinal mass, and risk stratification (intermediate/high risk). CONCLUSION: Our study was able to unveil TG level of >615 mg/dl, mediastinal mass, and a larger BSA as novel risk factors that have not been previously discussed in the literature.