Variational autoencoders (VAEs) are deep latent space generative models that have been immensely successful in multiple exciting applications in biomedical informatics such as molecular design, protein design, medical image classification and segmentation, integrated multi-omics data analyses, and large-scale biological sequence analyses, among others. The fundamental idea in VAEs is to learn the distribution of data in such a way that new meaningful data with more intra-class variations can be generated from the encoded distribution. The ability of VAEs to synthesize new data with more representation variance at state-of-art levels provides hope that the chronic scarcity of labeled data in the biomedical field can be resolved. Furthermore, VAEs have made nonlinear latent variable models tractable for modeling complex distributions. This has allowed for efficient extraction of relevant biomedical information from learned features for biological data sets, referred to as unsupervised feature representation learning. In this article, we review the various recent advancements in the development and application of VAEs for biomedical informatics. We discuss challenges and future opportunities for biomedical research with respect to VAEs.

Variational Auto-Encoders (VAEs) are deep latent space generative models which have been immensely successful in many applications such as image generation, image captioning, protein design, mu-tation prediction, and language models among others. The fundamental idea in VAEs is to learn the distribu-tion of data in such a way that new meaningful data can be generated from the encoded distribution. This concept has led to tremendous research and variations in the design of VAEs in the last few years creating a field of its own, referred to as unsupervised representation learning. This paper provides a much-needed comprehensive evaluation of the variations of the VAEs based on their end goals and resulting architectures. It further provides intuition as well as mathematical formulation and quantitative results of each popular variation, presents a concise comparison of these variations, and concludes with challenges and future oppor-tunities for research in VAEs.

Study Objectives: To use relatively noisy routinely collected clinical data (brain magnetic resonance imaging (MRI) data, clinical polysomnography (PSG) recordings, and neuropsychological testing), to investigate hypothesis-driven and data-driven relationships between brain physiology, structure, and cognition. Methods: We analyzed data from patients with clinical PSG, brain MRI, and neuropsychological evaluations. SynthSeg, a neural network-based tool, provided high-quality segmentations despite noise. A priori hypotheses explored associations between brain function (measured by PSG) and brain structure (measured by MRI). Associations with cognitive scores and dementia status were studied. An exploratory data-driven approach investigated age-structure-physiology-cognition links. Results: Six hundred and twenty-three patients with sleep PSG and brain MRI data were included in this study; 160 with cognitive evaluations. Three hundred and forty-two participants (55%) were female, and age interquartile range was 52 to 69 years. Thirty-six individuals were diagnosed with dementia, 71 with mild cognitive impairment, and 326 with major depression. One hundred and fifteen individuals were evaluated for insomnia and 138 participants had an apnea–hypopnea index equal to or greater than 15. Total PSG delta power correlated positively with frontal lobe/thalamic volumes, and sleep spindle density with thalamic volume. rapid eye movement (REM) duration and amygdala volume were positively associated with cognition. Patients with dementia showed significant differences in five brain structure volumes. REM duration, spindle, and slow-oscillation features had strong associations with cognition and brain structure volumes. PSG and MRI features in combination predicted chronological age (R2 = 0.67) and cognition (R2 = 0.40). Conclusions: Routine clinical data holds extended value in understanding and even clinically using brain-sleep-cognition relationships.

Despite the importance of few-shot learning, the lack of labeled training data in the real world makes it extremely challenging for existing machine learning methods because this limited dataset does not well represent the data variance. In this research, we suggest employing a generative approach using variational autoencoders (VAEs), which can be used specifically to optimize few-shot learning tasks by generating new samples with more intra-class variations on the Labeled Faces in the Wild (LFW) dataset. The purpose of our research is to increase the size of the training dataset using various methods to improve the accuracy and robustness of the few-shot face recognition. Specifically, we employ the VAE generator to increase the size of the training dataset, including the basic and the novel sets while utilizing transfer learning as the backend. Based on extensive experimental research, we analyze various data augmentation methods to observe how each method affects the accuracy of face recognition. The face generation method based on VAEs with perceptual loss can effectively improve the recognition accuracy rate to 96.47% using both the base and the novel sets.

Despite the importance of few-shot learning, the lack of labeled training data in the real world, makes it extremely challenging for existing machine learning methods as this limited data set does not represent the data variance well. In this research, we suggest employing a generative approach using variational autoencoders (VAEs), which can be used specifically to optimize few-shot learning tasks by generating new samples with more intra-class variations. The purpose of our research is to increase the size of the training data set using various methods to improve the accuracy and robustness of the few-shot face recognition. Specifically, we employ the VAE generator to increase the size of the training data set, including the basic and the novel sets while utilizing transfer learning as the backend. Based on extensive experimental research, we analyze various data augmentation methods to observe how each method affects the accuracy of face recognition. We conclude that the face generation method we proposed can effectively improve the recognition accuracy rate to 96.47% using both the base and the novel sets.

Background: Congestive heart failure (CHF) is a common cause of hospital admissions. Medical records contain valuable information about CHF, but manual chart review is time-consuming. Claims databases (using International Classification of Diseases [ICD] codes) provide a scalable alternative but are less accurate. Automated analysis of medical records through natural language processing (NLP) enables more efficient adjudication but has not yet been validated across multiple sites. Objective: We seek to accurately classify the diagnosis of CHF based on structured and unstructured data from each patient, including medications, ICD codes, and information extracted through NLP of notes left by providers, by comparing the effectiveness of several machine learning models. Methods: We developed an NLP model to identify CHF from medical records using electronic health records (EHRs) from two hospitals (Mass General Hospital and Beth Israel Deaconess Medical Center; from 2010 to 2023), with 2800 clinical visit notes from 1821 patients. We trained and compared the performance of logistic regression, random forests, and RoBERTa models. We measured model performance using area under the receiver operating characteristic curve (AUROC) and area under the precision-recall curve (AUPRC). These models were also externally validated by training the data on one hospital sample and testing on the other, and an overall estimated error was calculated using a completely random sample from both hospitals. Results: The average age of the patients was 66.7 (SD 17.2) years; 978 (54.3%) out of 1821 patients were female. The logistic regression model achieved the best performance using a combination of ICD codes, medications, and notes, with an AUROC of 0.968 (95% CI 0.940-0.982) and an AUPRC of 0.921 (95% CI 0.835-0.969). The models that only used ICD codes or medications had lower performance. The estimated overall error rate in a random EHR sample was 1.6%. The model also showed high external validity from training on Mass General Hospital data and testing on Beth Israel Deaconess Medical Center data (AUROC 0.927, 95% CI 0.908-0.944) and vice versa (AUROC 0.968, 95% CI 0.957-0.976). Conclusions: The proposed EHR-based phenotyping model for CHF achieved excellent performance, external validity, and generalization across two institutions. The model enables multiple downstream uses, paving the way for large-scale studies of CHF treatment effectiveness, comorbidities, outcomes, and mechanisms.