Autoencoder Research Articles

Unsupervised electrocardiogram feature extraction using deep learning empowers discovery of genetic and phenotypic determinants of cardiac electrophysiology

Abstract Background Conventional analysis of the electrocardiogram (ECG) is based on the extraction of human-defined, visually recognisable features. However, these are not optimised to capture all the information content in the ECG. The variational autoencoder (VAE), a form of unsupervised machine learning, can address this shortcoming by computationally extracting comprehensive and interpretable new ECG features, called latent factors. These latent factors provide a low-dimensional representation of the ECG that maximises capture of data content. This approach could expand our understanding of electrophysiology by identifying novel genetic and phenotypic traits associations with the ECG. Purpose This study aims to uncover genetic determinants of VAE latent factors, comparing them to determinants of conventional, human-derived ECG parameters to gain novel insights into cardiac electrophysiology and related diseases. Methods Over one million median beat ECGs derived from a United States (US) based secondary care centre were used to train a machine learning VAE model, with external validation in the UK Biobank (UKB). We performed common and rare variant association studies for VAE latent factors and conventional ECG traits on quality-controlled UKB data. Associated genetic variants were compared to loci for conventional ECG parameters available in the UKB and literature. Novel GWAS associations were validated in a withheld subset of the UKB cohort. Additionally, we compared the associations of the VAE latent factors and conventional ECG traits with phenotypic traits, disease codes and echocardiographic traits. Results Utilising median beat ECGs, the VAE model identified 20 features, with partial correlations to traditional ECG traits. A genome-wide association study (GWAS) identified 105 associated genomic regions, compared with 62 regions identified with conventional ECG traits on the same dataset. Most discovered loci have been previously associated with ECG features in larger GWAS, providing a positive control validation. We also discovered and validated five genomic regions not previously linked to the ECG (mapped to TRIOBP, EFEMP1, NEDD9, ATP10A and P2RX1). We found rare variant associations for seven genes (NEK6, IL17RA, NME7, MYBPC3, CCT8, ADAMTS6 and SCN5A). The latent factors uncovered associations with 158 phenotypic traits (2093 phenotypes tested) and 147 disease codes (2074 disease phecodes tested) which were not identified by ECG traits. Out of 38 echocardiographic traits, 35 showed higher correlations with the latent factors, while six exhibited significant associations exclusively with the latent factors. Conclusion Our study demonstrates that VAE-derived latent factors outperform conventional ECG parameters in uncovering novel genetic and phenotypic associations. Our findings underscore the opportunity for optimised data-driven feature extraction to enhance our understanding of cardiac electrophysiology.

Classification of Vaginal Cleanliness Grades through Surface‐Enhanced Raman Spectral Analysis via The Deep‐Learning Variational Autoencoder–Long Short‐Term Memory Model

In this study, it is aimed to establish a novel method based on a deep‐learning‐guided surface‐enhanced Raman spectroscopy (SERS) technique to achieve rapid and accurate classification of vaginal cleanliness levels. We proposed a variational autoencoder (VAE) approach to enhance spectral quality, coupled with a deep learning algorithm long short‐term memory (LSTM) neural network to analyze SERS spectra produced by vaginal secretions. The performance of various machine learning (ML) algorithms is assessed using multiple evaluation metrics. Finally, the reliability of the optimal model is tested using blind test data (N = 10/group for each cleanliness level). The data quality of the SERS fingerprints of four types of vaginal secretions is significantly improved after VAE decoding and reconstruction. The signal‐to‐noise ratio of the generated spectra increased from the original 2.58–11.13. Among all algorithms, the VAE–LSTM algorithm demonstrates the best prediction ability and time efficiency. Additionally, blind test datasets yielded an overall accuracy of 85%. In this study, it is concluded that the deep‐learning‐guided SERS technique holds significant potential in rapidly distinguishing between different levels of vaginal cleanliness through human vaginal secretion samples. This contributes to the efficient diagnosis of vaginal cleanliness levels in clinical settings.

Interpretable machine learning model for predicting long-term cardiovascular mortality risk in chronic kidney disease patients using SHAP explainers

Abstract Background Patients with chronic kidney disease (CKD) exhibited a pronounced burden of cardiac mortality. However, early prediction of cardiovascular mortality risk in CKD patients remains an unmet medical need. Purpose This study aimed to evaluate the performance of machine learning (ML) algorithms and identify the most influential factors in predicting long-term cardiac mortality among CKD patients. Methods Six ML models were developed, with the most effective one chosen for predicting and categorising patients into high-risk and low-risk groups based on the maximal Youden’s index. Differences in survival rates were assessed using the log-rank test on Kaplan-Meier curves. Cox regression analysis was employed to investigate the relationship between ML-predicted risk scores and mortality. Furthermore, the SHapley Additive exPlanations (SHAP) method was implemented to provide personalised explanations for model decisions. Results The auto-encoder (AE) model achieved the highest AUC of 0.94, with a sensitivity of 80.95% and specificity of 96.34%. Compared to individuals in low-risk group, those in high-risk group exhibited a significantly elevated risk of cardiovascular mortality (adjusted hazard ratio [HR]: 48.24; P <0.001). According to the SHAP model, age, C-reactive protein, blood urea nitrogen and hypertension were identified as the four most influential factors in the AE model. Conclusions Our study successfully developed an AE machine learning model for predicting 10-year cardiovascular mortality among CKD patients. The combination of ML model and SHAP aids doctors in understanding key features in the model, improving their insight into decision-making related to disease severity assessment, thus facilitating early preventive strategies in clinical practice.

Utilizing unsupervised machine learning to uncover novel phenogroups within the broad QRS complex

Abstract Background Conduction disease can manifest as a broad QRS complex on the ECG. Broad QRS is conventionally categorised as left bundle branch block (LBBB) and right bundle branch block (RBBB) or non-specific intraventricular conduction delay (IVCD), based on QRS morphology. These subgroups were coined over a century ago and have been relevant for heart failure with reduced ejection fraction and cardiac resynchronization therapy (CRT). However, there may be more precise phenogroups underlying broad QRS complexes. Purpose Using unsupervised machine learning to define novel broad QRS phenogroups and characterise them using clinical variables and disease outcomes. Methods We extracted 51 relevant features using a Variational Autoencoder from 47,456 median-beat ECGs with broad QRS intervals (QRS ≥ 130ms) from a secondary care cohort. These were input into Discriminative Dimensionality Reduction via Learning a Tree (DDRTree), a clustering method that models the continuum of heterogeneity by projecting the underlying distribution as a tree structure in a low-dimensional space, while assigning distinct branch/cluster labels. Results Six distinct phenogroups were identified within broad QRS morphologies (Figure-1). Two branches were RBBB dominant (Branch 1: more males, high-comorbidity; Branch 6: low-comorbidity/healthy-RBBB). Two branches were LBBB dominant (Branch 4: mixed IVCD with CHB burden; Branch 5: more females, CRT responders). Two branches were IVCD/mixed-phenotype (Branch 2: IVCD-dominant, high risk of future cardiovascular events; Branch 3: average/core branch). Within LBBB, Branch 4 and 5 showed similar trends with improving echocardiography profiles and CRT response when adjusted for covariates, each compared against the core branch. Branch 5 was distinct due its high risk of future HF (HR 1.19 [1.07-1.33] p < 0.01), cardiovascular death (HR 1.35 [1.20-1.50] p < 0.0001), all-cause mortality (HR 1.08 [1.01-1.16] p < 0.05) and CHB (HR 1.35 [1.16-1.56] p < 0.0001) (Figure-2). Conversely within RBBB, Branch 1 and 6 showed stark differences across various traits. Branch 1 captured a high-risk profile associated to higher risk of cardiovascular death (HR 1.37 [1.24-1.52] p < 0.0001), all-cause mortality (HR 1.20 [1.12-1.29] p < 0.0001) and CHB (HR 1.15 [1.01-1.32] p < 0.05) (Figure-2), while Branch 6 retained a low-risk phenotype. The tree dimensions also revealed relevant trends in cardiac anatomy and valvular dysfunction; the top-left region in Figure-1 associated with higher left ventricular ejection fraction and lower left ventricular end-diastolic dimension, lower left ventricular end-systolic dimension and a greater CRT response. Conclusion Through unsupervised methods we have identified conduction disease phenogroups with additive value for disease prognosis and CRT response prediction beyond traditional LBBB and RBBB labels. These novel phenogroups may help guide precision care for patients with a broad QRS complex.Figure-1Figure-2

Privacy-preserving MTS anomaly detection for network devices through federated learning

In the context of Maintenance-as-a-Service (MaaS), it is important for device vendors to develop multivariate time series (MTS) anomaly detection models that can accurately identify anomalies without compromising the privacy of customer enterprises' data. In this paper, we investigate the relationship between MTS data patterns and the parameters of unsupervised autoencoder (AE) models and show that they are highly consistent. Building on this insight, we propose a novel unsupervised federated learning (FL)-based framework called OmniFed, which cannot only address the heterogeneity of non-independent identically (non-iid) distributed data on different devices, but also achieve high-precision detection of device MTS anomalies while ensuring privacy. Specifically, OmniFed is initialized with an AE model and then trains local AE models on individual devices via federated learning. Finally, OmniFed clusters devices based on the parameters of the AE models and trains a cluster-specific MTS anomaly detection model using FL. Our experiments on two real-world datasets demonstrate that OmniFed achieves an F1-Score of 0.921, significantly higher than the best baseline method.

A Dataset of 3M Single-DOF Planar 4-, 6-, and 8-bar Linkage Mechanisms with Open and Closed Coupler Curves for Machine Learning-Driven Path Synthesis

Abstract In recent years, there has been a strong interest in applying machine learning techniques to path synthesis of linkage mechanisms. However, progress has been stymied due to a scarcity of high-quality datasets. In this paper, we present a comprehensive dataset comprising nearly three million samples of four-, six-, and eight-bar linkage mechanisms with open and closed coupler curves. Current machine learning approaches to path synthesis also lack standardized metrics for evaluating outcomes. To address this gap, we propose six key metrics to quantify results, providing a foundational framework for researchers to compare new models with existing ones. We also present a Variational AutoEncoder based model in conjunction with a k-nearest neighbor search approach to demonstrate the utility of our dataset. In the end, we provide example mechanisms that generate various curves along with numerical evaluation of the proposed metrics.

Performance comparison on improved data-driven building energy prediction under data shortage scenarios in four perspectives: Data generation, incremental learning, transfer learning, and physics-informed

Accurate building energy predictions (BEPs) are crucial for maintaining a built environment's sustainability and energy systems. Many data-driven BEPs rely heavily on sufficient data. However, circumstances of practical data shortage restrict the application of these data-driven models to information-poor buildings. This study aims to address this research gap by enhancing BEP performance via the consideration of four perspectives: data generation (DG), incremental learning (IL), transfer learning (TL) and physics-informed (PI). Long short-term memory (LSTM) is selected as the baseline model for comparing performance under extreme (target building dataset: 1-week), heavy (4-weeks) and mild (13-weeks) data shortage scenarios with respect to prediction errors, time costs and the Taylor diagram, among others. The implementation algorithms of four LSTM improved models are conditional variational autoencoder (CVAE), crude incremental learning, domain-adversarial neural network and statistics-informed (SI) for DG, IL, TL and PI, respectively. The effectiveness of the four enhanced models was confirmed using the open-source dataset Building Data Genome Project 2. A comprehensive and fair performance comparison recommends CVAE for extreme and heavy data shortage scenarios with average mean absolute per cent error (MAPE) of 12.79 % and 8.63 % as well as performance improvement ratios (PIRs) of 0.45 and 0.57. SI should be recommended for mild data shortage scenarios with MAPE of 8.01 % and PIR of 0.66.

Handling missing data in variational autoencoder based item response theory.

Recently Variational Autoencoders (VAEs) have been proposed as a method to estimate high dimensional Item Response Theory (IRT) models on large datasets. Although these improve the efficiency of estimation drastically compared to traditional methods, they have no natural way to deal with missing values. In this paper, we adapt three existing methods from the VAE literature to the IRT setting and propose one new method. We compare the performance of the different VAE-based methods to each other and to marginal maximum likelihood estimation for increasing levels of missing data in a simulation study for both three- and ten-dimensional IRT models. Additionally, we demonstrate the use of the VAE-based models on an existing algebra test dataset. Results confirm that VAE-based methods are a time-efficient alternative to marginal maximum likelihood, but that a larger number of importance-weighted samples are needed when the proportion of missing values is large.

Anomaly detection scheme for lung CT images using vector quantized variational auto-encoder with support vector data description.

This study aims to develop an anomaly-detection scheme for lesions in CT images. Our database consists of lung CT images obtained from 1500 examinees. It includes 1200 normal and 300 abnormal cases. In this study, SVDD (Support Vector Data Description) mapping the normal latent variables into a hypersphere as small as possible on the latent space is introduced to VQ-VAE (Vector Quantized-Variational Auto-Encoder). VQ-VAE with SVDD is constructed from two encoders, two decoders, and an embedding space. The first encoder compresses the input image into the latent-variable map, whereas the second encoder maps the normal latent variables into a hypersphere as small as possible. The first decoder then up-samples the mapped latent variables into a latent-variable map with the original size. The second decoder finally reconstructs the input image from the latent-variable map replaced by the embedding representations. The data of each examinee is classified as abnormal or normal based on the anomaly score defined as the combination of the difference between the input image and the reconstructed image and the distance between the latent variables and the center of the hypersphere. The area under the ROC curve for VQ-VAE with SVDD was 0.76, showing an improvement when compared with the conventional VAE (0.63, p < .001). VQ-VAE with SVDD developed in this study can yield higher anomaly-detection accuracy than the conventional VAE. The proposed method is expected to be useful for identifying examinees with lesions and reducing interpretation time in CT screening.

Empowering legal justice with AI: A reinforcement learning SAC-VAE framework for advanced legal text summarization.

Automated summarization of legal texts poses a significant challenge due to the complex and specialized nature of legal documentation. Despite the recent progress in reinforcement learning for natural language text summarization, its application in the legal domain has been less effective. This paper introduces SAC-VAE, a novel reinforcement learning framework specifically designed for legal text summarization. We leverage a Variational Autoencoder (VAE) to condense the high-dimensional state space into a more manageable lower-dimensional feature space. These compressed features are subsequently utilized by the Soft Actor-Critic (SAC) algorithm for policy learning, facilitating the automated generation of summaries from legal texts. Through comprehensive experimentation, we have empirically demonstrated the effectiveness and superior performance of the SAC-VAE framework in legal text summarization.

Understanding and mitigating the impact of race with adversarial autoencoders

BackgroundArtificial intelligence carries the risk of exacerbating some of our most challenging societal problems, but it also has the potential of mitigating and addressing these problems. The confounding effects of race on machine learning is an ongoing subject of research. This study aims to mitigate the impact of race on data-derived models, using an adversarial variational autoencoder (AVAE). In this study, race is a self-reported feature. Race is often excluded as an input variable, however, due to the high correlation between race and several other variables, race is still implicitly encoded in the data.MethodsWe propose building a model that (1) learns a low dimensionality latent spaces, (2) employs an adversarial training procedure that ensure its latent space does not encode race, and (3) contains necessary information for reconstructing the data. We train the autoencoder to ensure the latent space does not indirectly encode race.ResultsIn this study, AVAE successfully removes information about race from the latent space (AUC ROC = 0.5). In contrast, latent spaces constructed using other approaches still allow the reconstruction of race with high fidelity. The AVAE’s latent space does not encode race but conveys important information required to reconstruct the dataset. Furthermore, the AVAE’s latent space does not predict variables related to race (R2 = 0.003), while a model that includes race does (R2 = 0.08).ConclusionsThough we constructed a race-independent latent space, any variable could be similarly controlled. We expect AVAEs are one of many approaches that will be required to effectively manage and understand bias in ML.

AI-driven transcriptome profile-guided hit molecule generation

Denovo generation of bioactive and drug-like hit molecules is a pivotal goal in computer-aided drug discovery. While artificial intelligence (AI) has proven adept at generating molecules with desired chemical properties, previous studies often overlook the influence of disease-specific cellular environments. This study introduces GxVAEs, a novel AI-driven deep generative model designed to produce hit molecules from transcriptome profiles using dual variational autoencoders (VAEs). The first VAE, ProfileVAE, extracts latent features from transcriptome profiles to guide the second VAE, MolVAE, in generating hit molecules. GxVAEs aim to bridge the gap between molecule generation and the biological context of disease, producing molecules that are biologically relevant within specific cellular environments or pathological conditions. Experimental results and case studies focused on hit molecule generation demonstrate that GxVAEs surpass current state-of-the-art methods, in terms of reproducibility of known ligands. This approach is expected to effectively find potential molecular structures with bioactivities across diverse disease contexts.

Deep Learning‐Based Variational Autoencoder for Classification of Quantum and Classical States of Light

AbstractAdvancements in optical quantum technologies have been enabled by the generation, manipulation, and characterization of light, with identification based on its photon statistics. However, characterizing light and its sources through single photon measurements requires efficient detectors and longer measurement times to obtain high‐quality photon statistics. Here, a deep learning‐based variational autoencoder (VAE) method is introduced for classifying single photon added coherent state (SPACS), single photon added thermal state (SPATS), and mixed states between coherent and SPACS as well as between thermal and SPATS of light. The semi‐supervised learning‐based VAE efficiently maps the photon statistics features of light to a lower dimension, enabling quasi‐instantaneous classification with low average photon counts. The proposed VAE method is robust and maintains classification accuracy in the presence of losses inherent in an experiment, such as finite collection efficiency, non‐unity quantum efficiency, finite number of detectors, etc. Additionally, leveraging the transfer learning capabilities of VAE enables successful classification of data of any quality using a single trained model. It is envisioned that such a deep learning methodology will enable better classification of quantum light and light sources even in poor detection.

ReMiND: Recovery of missing neuroimaging using diffusion models with application to Alzheimer’s disease

Abstract Missing data is a significant challenge in medical research. In longitudinal studies of Alzheimer’s disease (AD) where structural magnetic resonance imaging (MRI) is collected from individuals at multiple time points, participants may miss a study visit or drop out. Additionally, technical issues such as participant motion in the scanner may result in unusable imaging data at designated visits. Such missing data may hinder the development of high-quality imaging-based biomarkers. To address the problem of missing MRI data in studies of AD, we introduced a novel 3D diffusion model specifically designed for imputing missing structural MRI (Recovery of Missing Neuroimaging using Diffusion models (ReMiND)). The model generates a whole-brain image conditional on a single structural MRI observed at a past visit or conditional on one past and one future observed structural MRI relative to the missing observation. The performance of models was compared with two alternative imputation approaches: forward filling and image generation using variational autoencoders. Experimental results show that our method can generate 3D structural MRI with high similarity to ground-truth images at designated visits. Furthermore, images generated using ReMiND show relatively lower differences in volume estimation between the imputed and observed images compared to images generated by forward filling or autoencoders. Additionally, ReMiND provides more accurate estimated rates of atrophy over time in important anatomical brain regions than the two comparator methods. Our 3D diffusion model can impute missing structural MRI data at a single designated visit and outperforms alternative methods for imputing whole-brain images that are missing from longitudinal trajectories.

Optimized beamforming vector and antenna selection using autoencoder with Bayesian RNN and hybrid heuristic algorithm for MIMO systems

SummaryWireless communication systems depend on channel state information (CSI). However, the CSI collection is a significant barrier to achieving high performance. Present wireless communication models mainly rely on many antennas and related signal dispensation for increasing effectiveness. User mobility, the usage of greater frequencies, and a better number of antennas contribute to beamforming and beam management in these networks. Therefore, an improved hybrid heuristic algorithm is proposed here for providing optimized Joint beam forming and antenna selection (JBAS) in a multi‐input multi‐output (MIMO) model. The major novelty of the research work is to jointly perform the beam forming and antenna selection in MIMO that effectively enhanced the performance of the wireless transmission system. The major intention of this paper is to optimize the beamforming vectors and antennae by using the designed algorithm, where the spectral efficiency gets maximized. Initially, the data is collected by varying the number of users, locations, and sampling points, which is optimally selected by developing the hybrid reptile search‐based lion optimization algorithm (HRS‐LOA). Once it is collected, the relevant CSI is considered. This state of information is fed as input to the Auto Encoder with Bayesian Recurrent Neural Network (AE‐BRNN). Finally, through the network, the optimized value of the beamforming vector and antenna is obtained, which is useful to increase the sum rate/spectral efficiency. From the numerical results, the recommended model attained a spectral efficiency of 30 bits/s/Hz, which is higher than the BWO‐AE‐BRNN with 13%, RSO‐AE‐BRNN with 16%, RSA‐AE‐BRNN with 17%, and LA‐AE‐BRNN with 40% respectively.

An efficient enhanced stacked auto encoder assisted optimized deep neural network for forecasting Dry Eye Disease

Meibomian Gland Dysfunction (MGD) and Dry Eye Disease (DED) comprise two of the most significant eye diseases, impacting millions of sufferers worldwide. Several etiological factors influence the early symptoms of DED. Early diagnosis and treatment of erectile dysfunction may significantly improve the Quality of Life (QoL) for people. The current study introduces the ESAE-ODNN, an improved stacked autoencoder-aided optimised deep neural network, as a new way to predict DED using feature selection (FS), feature extraction (FE), and classification. The approach described here is novel because it merges chaotic maps into FS, employs SLSTM-STSA for improved classification accuracy (CA), and optimizes with the adaptive quantum rotation of the Enhanced Quantum Bacterial Foraging Optimisation Algorithm (EQBFOA). The present study enhances prediction functions by extracting MGD-related features and complicated relationships from the DED dataset. To ensure essential feature identification, the ESAE minimizes irrelevant and redundant features. To predict the DED, the ESAE first applies FE and then implements an ODNN classifier. This method fine-tunes the ODNN framework to enhance the effectiveness of the classification. The proposed ESAE-ODNN classification system efficiently assists in the early diagnosis of DED. Combining advanced Deep Learning (DL) methods with optimization can help us understand MGD features better and sort the data with the best accuracy (96.34%). The experimental evaluation with relevant performance metrics indicates that the proposed method is efficient in diverse aspects: accurate identification, reduced complexity, and fine-tuned performance. The ESAE-ODNN’s robustness in handling intricate feature indications and high-dimensional data outperforms the existing state-of-the-art techniques.

CycleDiffusion: Voice Conversion Using Cycle-Consistent Diffusion Models

Voice conversion (VC) refers to the technique of modifying one speaker’s voice to mimic another’s while retaining the original linguistic content. This technology finds its applications in fields such as speech synthesis, accent modification, medicine, security, privacy, and entertainment. Among the various deep generative models used for voice conversion, including variational autoencoders (VAEs) and generative adversarial networks (GANs), diffusion models (DMs) have recently gained attention as promising methods due to their training stability and strong performance in data generation. Nevertheless, traditional DMs focus mainly on learning reconstruction paths like VAEs, rather than conversion paths as GANs do, thereby restricting the quality of the converted speech. To overcome this limitation and enhance voice conversion performance, we propose a cycle-consistent diffusion (CycleDiffusion) model, which comprises two DMs: one for converting the source speaker’s voice to the target speaker’s voice and the other for converting it back to the source speaker’s voice. By employing two DMs and enforcing a cycle consistency loss, the CycleDiffusion model effectively learns both reconstruction and conversion paths, producing high-quality converted speech. The effectiveness of the proposed model in voice conversion is validated through experiments using the VCTK (Voice Cloning Toolkit) dataset.

Reverse design and application of phononic crystals based on deep learning

Abstract This paper reverse-design phononic crystals with band gaps within a targeted frequency band using the trained Conditional Variational Autoencoder (CVAE) and further studies the vibro-acoustic characteristics of a composite sandwich plate with a phononic crystal panel as the core layer. Firstly, a matrix composed of 0s and 1s, representing scatterers and substrates, is randomly generated by MATLAB to represent two-dimensional phononic crystals. The three-dimensional phononic crystals are obtained by stretching the two-dimensional phononic crystals along the average direction, and COMSOL Multiphysics is used to calculate the band gap. In order to maximize the production of phononic crystals with a band gap distribution, the Convolutional Neural Network (CNN) is trained to predict whether the generated phononic crystals have band gaps. Finally, using data on the structures of phononic crystals and their band gap distributions, the CVAE is trained to achieve the reverse design of artificial periodic structures based on the target band gap. To verify the effectiveness of the structures obtained through the reverse design method on vibration and noise reduction, the submerged vibro-acoustic characteristics of a composite sandwich plate consisting of a phononic crystal panel and carbon fiber panels are studied. The model of the composite sandwich plate is fabricated, and its submerged vibro-acoustic characteristics are tested and compared with numerical results. Finally, the submerged vibro-acoustic response levels of composite sandwich plates with phononic crystal panels and honeycomb panels as core layers are compared using numerical methods to assess the phononic crystal panel's vibration and noise reduction effects.

Variational Autoencoding with Conditional Iterative Sampling for Missing Data Imputation

Variational autoencoders (VAEs) are popular for their robust nonlinear representation capabilities and have recently achieved notable advancements in the problem of missing data imputation. However, existing imputation methods often exhibit instability due to the inherent randomness in the sampling process, leading to either underestimation or overfitting, particularly when handling complex missing data types such as images. To address this challenge, we introduce a conditional iterative sampling imputation method. Initially, we employ an importance-weighted beta variational autoencoder to learn the conditional distribution from the observed data. Subsequently, leveraging the importance-weighted resampling strategy, samples are drawn iteratively from the conditional distribution to compute the conditional expectation of the missing data. The proposed method has been experimentally evaluated using classical generative datasets and compared with various well-known imputation methods to validate its effectiveness.

Leveraging the variational Bayes autoencoder for survival analysis

Survival analysis in medical research has witnessed a growing interest in applying deep learning techniques to model complex, high-dimensional, heterogeneous, incomplete, and censored data. Current methods make assumptions about the relations between data that may not be valid in practice. Therefore, we introduce SAVAE (Survival Analysis Variational Autoencoder). SAVAE, based on Variational Autoencoders, contributes significantly to the field by introducing a tailored Evidence Lower BOund formulation, supporting various parametric distributions for covariates and survival time (if the log-likelihood is differentiable). It offers a general method that demonstrates robustness and stability through different experiments. Our proposal effectively estimates time-to-event, accounting for censoring, covariate interactions, and time-varying risk associations. We validate our model in diverse datasets, including genomic, clinical, and demographic tabular data, with varying levels of censoring. This approach demonstrates competitive performance compared to state-of-the-art techniques, as assessed by the Concordance Index and the Integrated Brier Score. SAVAE also offers an interpretable model that parametrically models covariates and time. Moreover, its generative architecture facilitates further applications such as clustering, data imputation, and synthetic patient data generation through latent space inference from survival data. This approach fosters data sharing and collaboration, improving medical research and personalized patient care.