Prof. Dr.

Julia Vogt

Group Leader

E-Mail
julia.vogt@inf.ethz.ch
Phone
+41 44 633 8714
Address
Department of Computer Science
CAB G 69.1
Universitätstr. 6
CH – 8092 Zurich, Switzerland
Room
CAB G 69.1

Julia Vogt is an assistant professor in Computer Science at ETH Zurich, where she leads the Medical Data Science Group. The focus of her research is on linking computer science with medicine, with the ultimate aim of personalized patient treatment. She has studied mathematics both in Konstanz and in Sydney and earned her Ph.D. in computer science at the University of Basel. She was a postdoctoral research fellow at the Memorial Sloan-Kettering Cancer Center in NYC and with the Bioinformatics and Information Mining group at the University of Konstanz. In 2018, she joined the University of Basel as an assistant professor. In May 2019, she and her lab moved to Zurich where she joined the Computer Science Department of ETH Zurich.

Abstract

Survival analysis has gained significant attention in the medical domain and has many far-reaching applications. Although a variety of machine learning methods have been introduced for tackling time-to-event prediction in unstructured data with complex dependencies, clustering of survival data remains an under-explored problem. The latter is particularly helpful in discovering patient subpopulations whose survival is regulated by different generative mechanisms, a critical problem in precision medicine. To this end, we introduce a novel probabilistic approach to cluster survival data in a variational deep clustering setting. Our proposed method employs a deep generative model to uncover the underlying distribution of both the explanatory variables and the potentially censored survival times. We compare our model to the related work on survival clustering in comprehensive experiments on a range of synthetic, semi-synthetic, and real-world datasets. Our proposed method performs better at identifying clusters and is competitive at predicting survival times in terms of the concordance index and relative absolute error. To further demonstrate the usefulness of our approach, we show that our method identifies meaningful clusters from an observational cohort of hemodialysis patients that are consistent with previous clinical findings.

Authors

Laura Manduchi, Ricards Marcinkevics, Michela C. Massi, Verena Gotta, Timothy Müller, Flavio Vasella, Marian C. Neidert, Marc Pfister, Julia E. Vogt

Submitted

Arxiv

LinkCode

Abstract

Constrained clustering has gained significant attention in the field of machine learning as it can leverage prior information on a growing amount of only partially labeled data. Following recent advances in deep generative models, we propose a novel framework for constrained clustering that is intuitive, interpretable, and can be trained efficiently in the framework of stochastic gradient variational inference. By explicitly integrating domain knowledge in the form of probabilistic relations, our proposed model (DC-GMM) uncovers the underlying distribution of data conditioned on prior clustering preferences, expressed as pairwise constraints. These constraints guide the clustering process towards a desirable partition of the data by indicating which samples should or should not belong to the same cluster. We provide extensive experiments to demonstrate that DC-GMM shows superior clustering performances and robustness compared to state-of-the-art deep constrained clustering methods on a wide range of data sets. We further demonstrate the usefulness of our approach on two challenging real-world applications.

Authors

Laura Manduchi, Kieran Chin-Cheong, Holger Michel, Sven Wellmann, Julia E. Vogt

Submitted

arXiv

Link

Abstract

Background Preterm neonates frequently experience hypernatremia (plasma sodium concentrations >145 mmol/l), which is associated with clinical complications, such as intraventricular hemorrhage. Study design In this single center retrospective observational study, the following 7 risk factors for hypernatremia were analyzed in very low gestational age (VLGA, below 32 weeks) neonates: gestational age (GA), delivery mode (DM; vaginal or caesarian section), sex, birth weight, small for GA, multiple birth, and antenatal corticosteroids. Machine learning (ML) approaches were applied to obtain probabilities for hypernatremia. Results 824 VLGA neonates were included (median GA 29.4 weeks, median birth weight 1170g, caesarean section 83%). 38% of neonates experienced hypernatremia. Maximal sodium concentration of 144 mmol/l (interquartile range 142–147) was observed 52 hours (41–65) after birth. ML identified vaginal delivery and GA as key risk factors for hypernatremia. The risk of hypernatremia increased with lower GA from 22% for GA >= 31–32 weeks to 46% for GA < 31 weeks and 60% for GA < 27 weeks. A linear relationship between maximal sodium concentrations and GA was found, showing decreases of 0.29 mmol/l per increasing week GA in neonates with vaginal delivery and 0.49 mmol/l/week after cesarean section. Sex, multiple birth and antenatal corticosteroids were not associated hypernatremia. Conclusion VLGA neonates with vaginal delivery and low GA have the highest risk for hypernatremia. Early identification of neonates at risk and early intervention may prevent extreme sodium excursions and associated clinical complications.

Authors

Nadia S. Eugster, Florence Corminboeuf, Gilbert Koch, Julia E. Vogt, Thomas Sutter, Tamara van Donge, Marc Pfister, Roland Gerull

Submitted

Klinische Pädiatrie

LinkDOI

Abstract

Multiple data types naturally co-occur when describing real-world phenomena and learning from them is a long-standing goal in machine learning research. However, existing self-supervised generative models approximating an ELBO are not able to fulfill all desired requirements of multimodal models: their posterior approximation functions lead to a trade-off between the semantic coherence and the ability to learn the joint data distribution. We propose a new, generalized ELBO formulation for multimodal data that overcomes these limitations. The new objective encompasses two previous methods as special cases and combines their benefits without compromises. In extensive experiments, we demonstrate the advantage of the proposed method compared to state-of-the-art models in self-supervised, generative learning tasks.

Authors

Thomas M. Sutter, Imant Daunhawer, Julia E. Vogt

Submitted

Ninth International Conference on Learning Representations, ICLR 2021

Link

Abstract

Background: Given the absence of consolidated and standardized international guidelines for managing pediatric appendicitis and the few strictly data-driven studies in this specific, we investigated the use of machine learning (ML) classifiers for predicting the diagnosis, management and severity of appendicitis in children. Materials and Methods: Predictive models were developed and validated on a dataset acquired from 430 children and adolescents aged 0-18 years, based on a range of information encompassing history, clinical examination, laboratory parameters, and abdominal ultrasonography. Logistic regression, random forests, and gradient boosting machines were used for predicting the three target variables. Results: A random forest classifier achieved areas under the precision-recall curve of 0.94, 0.92, and 0.70, respectively, for the diagnosis, management, and severity of appendicitis. We identified smaller subsets of 6, 17, and 18 predictors for each of targets that sufficed to achieve the same performance as the model based on the full set of 38 variables. We used these findings to develop the user-friendly online Appendicitis Prediction Tool for children with suspected appendicitis. Discussion: This pilot study considered the most extensive set of predictor and target variables to date and is the first to simultaneously predict all three targets in children: diagnosis, management, and severity. Moreover, this study presents the first ML model for appendicitis that was deployed as an open access easy-to-use online tool. Conclusion: ML algorithms help to overcome the diagnostic and management challenges posed by appendicitis in children and pave the way toward a more personalized approach to medical decision-making. Further validation studies are needed to develop a finished clinical decision support system.

Authors

Ricards Marcinkevics, Patricia Reis Wolfertstetter, Sven Wellmann, Christian Knorr, Julia E Vogt

Submitted

Frontiers in Pediatrics

LinkDOI

Abstract

Survival analysis has gained significant attention in the medical domain with many far-reaching applications. Although a variety of machine learning methods have been introduced for tackling time-to-event prediction in unstructured data with complex dependencies, clustering of survival data remains an under-explored problem. The latter is particularly helpful in discovering patient subpopulations whose survival is regulated by different generative mechanisms, a critical problem in precision medicine. To this end, we introduce a novel probabilistic approach to cluster survival data in a variational deep clustering setting. Our proposed method employs a deep generative model to uncover the underlying distribution of both the explanatory variables and the potentially censored survival times. We compare our model to the related work on survival clustering in comprehensive experiments on a range of synthetic, semi-synthetic, and real-world datasets. Our proposed method performs better at identifying clusters and is competitive at predicting survival times in terms of the concordance index and relative absolute error.

Authors

Laura Manduchi, Ricards Marcinkevics, Julia E. Vogt

Submitted

Contributed talk at AI for Public Health Workshop at ICLR 2021

Link

Abstract

Generating interpretable visualizations of multivariate time series in the intensive care unit is of great practical importance. Clinicians seek to condense complex clinical observations into intuitively understandable critical illness patterns, like failures of different organ systems. They would greatly benefit from a low-dimensional representation in which the trajectories of the patients’ pathology become apparent and relevant health features are highlighted. To this end, we propose to use the latent topological structure of Self-Organizing Maps (SOMs) to achieve an interpretable latent representation of ICU time series and combine it with recent advances in deep clustering. Specifically, we (a) present a novel way to fit SOMs with probabilistic cluster assignments (PSOM), (b) propose a new deep architecture for probabilistic clustering (DPSOM) using a VAE, and (c) extend our architecture to cluster and forecastclinical states in time series (T-DPSOM). We show that our model achieves superior clustering performance compared to state-of-the-art SOM-based clustering methods while maintaining the favorable visualization properties of SOMs. On the eICU data-set, we demonstrate that T-DPSOM provides interpretable visualizations ofpatient state trajectories and uncertainty estimation. We show that our method rediscovers well-known clinical patient characteristics, such as a dynamic variant of the Acute Physiology And Chronic Health Evaluation (APACHE) score. Moreover, we illustrate how itcan disentangle individual organ dysfunctions on disjoint regions of the two-dimensional SOM map.

Authors

Laura Manduchi, Matthias Hüser, Martin Faltys, Julia Vogt, Gunnar Rätsch, Vincent Fortuin

Submitted

ACM CHIL 2021

Link

Abstract

In the quest for efficient and robust learning methods, combining unsupervised state representation learning and reinforcement learning (RL) could offer advantages for scaling RL algorithms by providing the models with a useful inductive bias. For achieving this, an encoder is trained in an unsupervised manner with two state representation methods, a variational autoencoder and a contrastive estimator. The learned features are then fed to the actor-critic RL algorithm Proximal Policy Optimization (PPO) to learn a policy for playing Open AI’s car racing environment. Hence, such procedure permits to decouple state representations from RL-controllers. For the integration of RL with unsupervised learning, we explore various designs for variational autoencoders and contrastive learning. The proposed method is compared to a deep network trained directly on pixel inputs with PPO. The results show that the proposed method performs slightly worse than directly learning from pixel inputs; however, it has a more stable learning curve, a substantial reduction of the buffer size, and requires optimizing 88% fewer parameters. These results indicate that the use of pre-trained state representations hasseveral benefits for solving RL tasks.

Authors

Juan M. Montoya, Imant Daunhawer, Julia E. Vogt, Marco Wiering

Submitted

ICAART 2021

Link

Abstract

Exploratory analysis of time series data can yield a better understanding of complex dynamical systems. Granger causality is a practical framework for analysing interactions in sequential data, applied in a wide range of domains. In this paper, we propose a novel framework for inferring multivariate Granger causality under nonlinear dynamics based on an extension of self-explaining neural networks. This framework is more interpretable than other neural-network-based techniques for inferring Granger causality, since in addition to relational inference, it also allows detecting signs of Granger-causal effects and inspecting their variability over time. In comprehensive experiments on simulated data, we show that our framework performs on par with several powerful baseline methods at inferring Granger causality and that it achieves better performance at inferring interaction signs. The results suggest that our framework is a viable and more interpretable alternative to sparse-input neural networks for inferring Granger causality.

Authors

Ricards Marcinkevics, Julia E. Vogt

Submitted

Ninth International Conference on Learning Representations, ICLR 2021

Link

Abstract

Rationale Tuberculosis diagnosis in children remains challenging. Microbiological confirmation of tuberculosis disease is often lacking, and standard immunodiagnostic including the tuberculin skin test and interferon-gamma release assay for tuberculosis infection has limited sensitivity. Recent research suggests that inclusion of novel Mycobacterium tuberculosis antigens has the potential to improve standard immunodiagnostic tests for tuberculosis. Objective To identify optimal antigen–cytokine combinations using novel Mycobacterium tuberculosis antigens and cytokine read-outs by machine learning algorithms to improve immunodiagnostic assays for tuberculosis. Methods A total of 80 children undergoing investigation of tuberculosis were included (15 confirmed tuberculosis disease, five unconfirmed tuberculosis disease, 28 tuberculosis infection and 32 unlikely tuberculosis). Whole blood was stimulated with 10 novel Mycobacterium tuberculosis antigens and a fusion protein of early secretory antigenic target (ESAT)-6 and culture filtrate protein (CFP) 10. Cytokines were measured using xMAP multiplex assays. Machine learning algorithms defined a discriminative classifier with performance measured using area under the receiver operating characteristics. Measurements and main results We found the following four antigen–cytokine pairs had a higher weight in the discriminative classifier compared to the standard ESAT-6/CFP-10-induced interferon-gamma: Rv2346/47c- and Rv3614/15c-induced interferon-gamma inducible protein-10; Rv2031c-induced granulocyte-macrophage colony-stimulating factor and ESAT-6/CFP-10-induced tumor necrosis factor-alpha. A combination of the 10 best antigen–cytokine pairs resulted in area under the curve of 0.92 +/- 0.04. Conclusion We exploited the use of machine learning algorithms as a key tool to evaluate large immunological datasets. This identified several antigen–cytokine pairs with the potential to improve immunodiagnostic tests for tuberculosis in children.

Authors

Noemi Rebecca Meier, Thomas M. Sutter, Marc Jacobsen, Tom H. M. Ottenhoff, Julia E. Vogt, Nicole Ritz

Submitted

Frontiers in Cellular and Infection Microbiology

LinkDOI

Abstract

Unplanned hospital readmissions are a burden to patients and increase healthcare costs. A wide variety of machine learning (ML) models have been suggested to predict unplanned hospital readmissions. These ML models were often specifically trained on patient populations with certain diseases. However, it is unclear whether these specialized ML models—trained on patient subpopulations with certain diseases or defined by other clinical characteristics—are more accurate than a general ML model trained on an unrestricted hospital cohort. In this study based on an electronic health record cohort of consecutive inpatient cases of a single tertiary care center, we demonstrate that accurate prediction of hospital readmissions may be obtained by general, disease-independent, ML models. This general approach may substantially decrease the cost of development and deployment of respective ML models in daily clinical routine, as all predictions are obtained by the use of a single model.

Authors

Thomas Sutter, Jan A Roth, Kieran Chin-Cheong, Balthasar L Hug, Julia E Vogt

Submitted

Journal of the American Medical Informatics Association

LinkDOI

Abstract

In this review, we examine the problem of designing interpretable and explainable machine learning models. Interpretability and explainability lie at the core of many machine learning and statistical applications in medicine, economics, law, and natural sciences. Although interpretability and explainability have escaped a clear universal definition, many techniques motivated by these properties have been developed over the recent 30 years with the focus currently shifting towards deep learning methods. In this review, we emphasise the divide between interpretability and explainability and illustrate these two different research directions with concrete examples of the state-of-the-art. The review is intended for a general machine learning audience with interest in exploring the problems of interpretation and explanation beyond logistic regression or random forest variable importance. This work is not an exhaustive literature survey, but rather a primer focusing selectively on certain lines of research which the authors found interesting or informative.

Authors

Ricards Marcinkevics, Julia E. Vogt

Submitted

Arxiv

Link

Abstract

Exploratory analysis of time series data can yield a better understanding of complex dynamical systems. Granger causality is a practical framework for analysing interactions in sequential data, applied in a wide range of domains. In this paper, we propose a novel framework for inferring multivariate Granger causality under nonlinear dynamics based on an extension of self-explaining neural networks. This framework is more interpretable than other neural-network-based techniques for inferring Granger causality, since in addition to relational inference, it also allows detecting signs of Granger-causal effects and inspecting their variability over time. In comprehensive experiments on simulated data, we show that our framework performs on par with several powerful baseline methods at inferring Granger causality and that it achieves better performance at inferring interaction signs. The results suggest that our framework is a viable and more interpretable alternative to sparse-input neural networks for inferring Granger causality.

Authors

Ricards Marcinkevics, Julia E. Vogt

Submitted

Interpretable Inductive Biases and Physically Structured Learning Workshop, NeurIPS 2020

Link

Abstract

Learning from different data types is a long-standing goal in machine learning research, as multiple information sources co-occur when describing natural phenomena. However, existing generative models that approximate a multimodal ELBO rely on difficult or inefficient training schemes to learn a joint distribution and the dependencies between modalities. In this work, we propose a novel, efficient objective function that utilizes the Jensen-Shannon divergence for multiple distributions. It simultaneously approximates the unimodal and joint multimodal posteriors directly via a dynamic prior. In addition, we theoretically prove that the new multimodal JS-divergence (mmJSD) objective optimizes an ELBO. In extensive experiments, we demonstrate the advantage of the proposed mmJSD model compared to previous work in unsupervised, generative learning tasks.

Authors

Thomas M. Sutter, Imant Daunhawer, Julia E. Vogt

Submitted

NeurIPS

Link

Abstract

PET/CT imaging is the gold standard for the diagnosis and staging of lung cancer. However, especially in healthcare systems with limited resources, costly PET/CT images are often not readily available. Conventional machine learning models either process CT or PET/CT images but not both. Models designed for PET/CT images are hence restricted by the number of PET images, such that they are unable to additionally leverage CT-only data. In this work, we apply the concept of visual soft attention to efficiently learn a model for lung cancer segmentation from only a small fraction of PET/CT scans and a larger pool of CT-only scans. We show that our model is capable of jointly processing PET/CT as well as CT-only images, which performs on par with the respective baselines whether or not PET images are available at test time. We then demonstrate that the model learns efficiently from only a few PET/CT scans in a setting where mostly CT-only data is available, unlike conventional models.

Authors

Varaha Karthik Pattisapu, Imant Daunhawer, Thomas Weikert, Alexander Sauter, Bram Stieltjes, Julia E. Vogt

Submitted

GCPR

Link

Abstract

Multimodal generative models learn a joint distribution over multiple modalities and thus have the potential to learn richer representations than unimodal models. However, current approaches are either inefficient in dealing with more than two modalities or fail to capture both modality-specific and shared variations. We introduce a new multimodal generative model that integrates both modality-specific and shared factors and aggregates shared information across any subset of modalities efficiently. Our method partitions the latent space into disjoint subspaces for modality-specific and shared factors and learns to disentangle these in a purely self-supervised manner. In extensive experiments, we show improvements in representation learning and generative performance compared to previous methods and showcase the disentanglement capabilities.

Authors

Imant Daunhawer, Thomas M. Sutter, Ricards Marcinkevics, Julia E. Vogt

Submitted

GCPR

Link

Abstract

Background The mortality risk remains significant in paediatric and adult patients on chronic haemodialysis (HD) treatment. We aimed to identify factors associated with mortality in patients who started HD as children and continued HD as adults. Methods The data originated from a cohort of patients < 30 years of age who started HD in childhood (<= 19 years) on thrice-weekly HD in outpatient DaVita dialysis centres between 2004 and 2016. Patients with at least 5 years of follow-up since the initiation of HD or death within 5 years were included; 105 variables relating to demographics, HD treatment and laboratory measurements were evaluated as predictors of 5-year mortality utilizing a machine learning approach (random forest). Results A total of 363 patients were included in the analysis, with 84 patients having started HD at < 12 years of age. Low albumin and elevated lactate dehydrogenase (LDH) were the two most important predictors of 5-year mortality. Other predictors included elevated red blood cell distribution width or blood pressure and decreased red blood cell count, haemoglobin, albumin:globulin ratio, ultrafiltration rate, z-score weight for age or single-pool K_t/V (below target). Mortality was predicted with an accuracy of 81%. Conclusions Mortality in paediatric and young adult patients on chronic HD is associated with multifactorial markers of nutrition, inflammation, anaemia and dialysis dose. This highlights the importance of multimodal intervention strategies besides adequate HD treatment as determined by K_t/V alone. The association with elevated LDH was not previously reported and may indicate the relevance of blood–membrane interactions, organ malperfusion or haematologic and metabolic changes during maintenance HD in this population.

Authors

Verena Gotta, Georgi Tancev, Olivera Marsenic, Julia E. Vogt, Marc Pfister

Submitted

Nephrology Dialysis Transplantation

LinkDOI

Abstract

Electronic Health Records (EHRs) are commonly used by the machine learning community for research on problems specifically related to health care and medicine. EHRs have the advantages that they can be easily distributed and contain many features useful for e.g. classification problems. What makes EHR data sets different from typical machine learning data sets is that they are often very sparse, due to their high dimensionality, and often contain heterogeneous (mixed) data types. Furthermore, the data sets deal with sensitive information, which limits the distribution of any models learned using them, due to privacy concerns. For these reasons, using EHR data in practice presents a real challenge. In this work, we explore using Generative Adversarial Networks to generate synthetic, heterogeneous EHRs with the goal of using these synthetic records in place of existing data sets for downstream classification tasks. We will further explore applying differential privacy (DP) preserving optimization in order to produce DP synthetic EHR data sets, which provide rigorous privacy guarantees, and are therefore shareable and usable in the real world. The performance (measured by AUROC, AUPRC and accuracy) of our model's synthetic, heterogeneous data is very close to the original data set (within 3 - 5% of the baseline) for the non-DP model when tested in a binary classification task. Using strong (1,10^-5) DP, our model still produces data useful for machine learning tasks, albeit incurring a roughly 17% performance penalty in our tested classification task. We additionally perform a sub-population analysis and find that our model does not introduce any bias into the synthetic EHR data compared to the baseline in either male/female populations, or the 0-18, 19-50 and 51+ age groups in terms of classification performance for either the non-DP or DP variant.

Authors

Kieran Chin-Cheong, Thomas M. Sutter, Julia E. Vogt

Submitted

Arxiv

Link

Abstract

Clinical pharmacology is a multi-disciplinary data sciences field that utilizes mathematical and statistical methods to generate maximal knowledge from data. Pharmacometrics (PMX) is a well-recognized tool to characterize disease progression, pharmacokinetics and risk factors. Since the amount of data produced keeps growing with increasing pace, the computational effort necessary for PMX models is also increasing. Additionally, computationally efficient methods such as machine learning (ML) are becoming increasingly important in medicine. However, ML is currently not an integrated part of PMX, for various reasons. The goals of this article are to (i) provide an introduction to ML classification methods, (ii) provide examples for a ML classification analysis to identify covariates based on specific research questions, (iii) examine a clinically relevant example to investigate possible relationships of ML and PMX, and (iv) present a summary of ML and PMX tasks to develop clinical decision support tools.

Authors

Gilbert Koch, Marc Pfister, Imant Daunhawer, Melanie Wilbaux, Sven Wellmann, Julia E. Vogt

Submitted

Clinical Pharmacology & Therapeutics

LinkDOI

Abstract

Despite the application of advanced statistical and pharmacometric approaches to pediatric trial data, a large pediatric evidence gap still remains. Here, we discuss how to collect more data from children by using real-world data from electronic health records, mobile applications, wearables, and social media. The large datasets collected with these approaches enable, and may demand, the use of artificial intelligence and machine learning to allow the data to be analyzed for decision-making. Applications of this approach are presented, which include the prediction of future clinical complications, medical image analysis, identification of new pediatric endpoints and biomarkers, the prediction of treatment non-responders and the prediction of placebo-responders for trial enrichment. Finally, we discuss how to bring machine learning from science to pediatric clinical practice. We conclude that advantage should be taken of the current opportunities offered by innovations in data science and machine learning to close the pediatric evidence gap.

Authors

Sebastiaan C. Goulooze, Laura B. Zwep, Julia E. Vogt, Elke H.J. Krekels, Thomas Hankemeier, John N. van den Anker, Catherijne A.J. Knibbe

Submitted

Clinical Pharmacology & Therapeutics

LinkDOI

Abstract

Learning from different data types is a long standing goal in machine learning research, as multiple information sources co-occur when describing natural phenomena. Existing generative models that try to approximate a multimodal ELBO rely on difficult training schemes to handle the intermodality dependencies, as well as the approximation of the joint representation in case of missing data. In this work, we propose an ELBO for multimodal data which learns the unimodal and joint multimodal posterior approximation functions directly via a dynamic prior. We show that this ELBO is directly derived from a variational inference setting for multiple data types, resulting in a divergence term which is the Jensen-Shannon divergence for multiple distributions. We compare the proposed multimodal JS-divergence (mmJSD) model to state-of-the-art methods and show promising results using our model in unsupervised, generative learning using a multimodal VAE on two different datasets.

Authors

Thomas Sutter, Imant Daunhawer, Julia E. Vogt

Submitted

Visually Grounded Interaction and Language Workshop, NeurIPS 2019

Abstract

Multimodal generative models learn a joint distribution of data from different modalities---a task which arguably benefits from the disentanglement of modality-specific and modality-invariant information. We propose a factorized latent variable model that learns named disentanglement on multimodal data without additional supervision. We demonstrate the disentanglement capabilities on simulated data, and show that disentangled representations can improve the conditional generation of missing modalities without sacrificing unconditional generation.

Authors

Imant Daunhawer, Thomas Sutter, Julia E. Vogt

Submitted

Bayesian Deep Learning Workshop, NeurIPS 2019

Abstract

Electronic Health Records (EHRs) are commonly used by the machine learning community for research on problems specifically related to health care and medicine. EHRs have the advantages that they can be easily distributed and contain many features useful for e.g. classification problems. What makes EHR data sets different from typical machine learning data sets is that they are often very sparse, due to their high dimensionality, and often contain heterogeneous data types. Furthermore, the data sets deal with sensitive information, which limits the distribution of any models learned using them, due to privacy concerns. In this work, we explore using Generative Adversarial Networks to generate synthetic, \textit{heterogeneous} EHRs with the goal of using these synthetic records in place of existing data sets. We will further explore applying differential privacy (DP) preserving optimization in order to produce differentially private synthetic EHR data sets, which provide rigorous privacy guarantees, and are therefore more easily shareable. The performance of our model's synthetic, heterogeneous data is very close to the original data set (within 4.5%) for the non-DP model. Although around 20% worse, the DP synthetic data is still usable for machine learning tasks.

Authors

Kieran Chin-Cheong, Thomas Sutter, Julia E. Vogt

Submitted

Machine Learning for Health (ML4H) Workshop, NeurIPS 2019

Abstract

We present a probabilistic model for clustering which enables the modeling of overlapping clusters where objects are only available as pairwise distances. Examples of such distance data are genomic string alignments, or protein contact maps. In our clustering model, an object has the freedom to belong to one or more clusters at the same time. By using an IBP process prior, there is no need to explicitly fix the number of clusters, as well as the number of overlapping clusters, in advance. In this paper, we demonstrate the utility of our model using distance data obtained from HIV1 protease inhibitor contact maps.

Authors

Sandhya Prabhakaran, Julia E. Vogt

Submitted

Artificial Intelligence in Medicine (AIME), Springer Lecture Notes in Artificial Intelligence, 2019

LinkDOI

Abstract

The recent adoption of Electronic Health Records (EHRs) by health care providers has introduced an important source of data that provides detailed and highly specific insights into patient phenotypes over large cohorts. These datasets, in combination with machine learning and statistical approaches, generate new opportunities for research and clinical care. However, many methods require the patient representations to be in structured formats, while the information in the EHR is often locked in unstructured texts designed for human readability. In this work, we develop the methodology to automatically extract clinical features from clinical narratives from large EHR corpora without the need for prior knowledge. We consider medical terms and sentences appearing in clinical narratives as atomic information units. We propose an efficient clustering strategy suitable for the analysis of large text corpora and to utilize the clusters to represent information about the patient compactly. To demonstrate the utility of our approach, we perform an association study of clinical features with somatic mutation profiles from 4,007 cancer patients and their tumors. We apply the proposed algorithm to a dataset consisting of about 65 thousand documents with a total of about 3.2 million sentences. We identify 341 significant statistical associations between the presence of somatic mutations and clinical features. We annotated these associations according to their novelty, and report several known associations. We also propose 32 testable hypotheses where the underlying biological mechanism does not appear to be known but plausible. These results illustrate that the automated discovery of clinical features is possible and the joint analysis of clinical and genetic datasets can generate appealing new hypotheses.

Authors

Stefan G. Stark, Stephanie L. Hyland, Melanie F. Pradier, Kjong Lehmann, Andreas Wicki, Fernando Perez Cruz, Julia E. Vogt, Gunnar Rätsch

Submitted

Arxiv preprint

Link

Abstract

Motivation: Personalized medicine aims at combining genetic, clinical, and environmental data to improve medical diagnosis and disease treatment, tailored to each patient. This paper presents a Bayesian nonparametric (BNP) approach to identify genetic associations with clinical/environmental features in cancer. We propose an unsupervised approach to generate data-driven hypotheses and bring potentially novel insights about cancer biology. Our model combines somatic mutation information at gene-level with features extracted from the Electronic Health Record. We propose a hierarchical approach, the hierarchical Poisson factor analysis (H-PFA) model, to share information across patients having different types of cancer. To discover statistically significant associations, we combine Bayesian modeling with bootstrapping techniques and correct for multiple hypothesis testing. Results: Using our approach, we empirically demonstrate that we can recover well-known associations in cancer literature. We compare the results of H-PFA with two other classical methods in the field: case-control (CC) setups, and linear mixed models (LMMs).

Authors

Melanie F. Pradier, Stephanie L. Hyland, Stefan G. Stark, Kjong Lehmann, Julia E. Vogt, Fernando Perez-Cruz, Gunnar Rätsch

Submitted

Biorxiv preprint

LinkDOI

Abstract

Background Machine learning models may enhance the early detection of clinically relevant hyperbilirubinemia based on patient information available in every hospital. Methods We conducted a longitudinal study on preterm and term born neonates with serial measurements of total serum bilirubin in the first two weeks of life. An ensemble, that combines a logistic regression with a random forest classifier, was trained to discriminate between the two classes phototherapy treatment vs. no treatment. Results Of 362 neonates included in this study, 98 had a phototherapy treatment, which our model was able to predict up to 48 h in advance with an area under the ROC-curve of 95.20%. From a set of 44 variables, including potential laboratory and clinical confounders, a subset of just four (bilirubin, weight, gestational age, hours since birth) suffices for a strong predictive performance. The resulting early phototherapy prediction tool (EPPT) is provided as an open web application. Conclusion Early detection of clinically relevant hyperbilirubinemia can be enhanced by the application of machine learning. Existing guidelines can be further improved to optimize timing of bilirubin measurements to avoid toxic hyperbilirubinemia in high-risk patients while minimizing unneeded measurements in neonates who are at low risk.

Authors

Imant Daunhawer, Severin Kasser, Gilbert Koch, Lea Sieber, Hatice Cakal, Janina Tütsch, Marc Pfister, Sven Wellman, Julia E. Vogt

Submitted

Pediatric Research, 2019

LinkDOI

Abstract

To exploit the full potential of big routine data in healthcare and to efficiently communicate and collaborate with information technology specialists and data analysts, healthcare epidemiologists should have some knowledge of large-scale analysis techniques, particularly about machine learning. This review focuses on the broad area of machine learning and its first applications in the emerging field of digital healthcare epidemiology.

Authors

Jan A. Roth, Manuel Battegay, Fabrice Juchler, Julia E. Vogt, Andreas F. Widmer

Submitted

Infection Control & Hospital Epidemiology, 2018

LinkDOI

Abstract

Molecular classification of hepatocellular carcinomas (HCC) could guide patient stratification for personalizedtherapies targeting subclass-specific cancer 'driver pathways'. Currently, there are several transcriptome-basedmolecular classifications of HCC with different subclass numbers, ranging from two to six. They were estab-lished using resected tumours that introduce a selection bias towards patients without liver cirrhosis and withearly stage HCCs. We generated and analyzed gene expression data from paired HCC and non-cancerous livertissue biopsies from 60 patients as well as five normal liver samples. Unbiased consensus clustering of HCCbiopsy profiles identified 3 robust classes. Class membership correlated with survival, tumour size and withEdmondson and Barcelona Clinical Liver Cancer (BCLC) stage. When focusing only on the gene expression ofthe HCC biopsies, we could validate previously reported classifications of HCC based on expression patterns ofsignature genes. However, the subclass-specific gene expression patterns were no longer preserved when thefold-change relative to the normal tissue was used. The majority of genes believed to be subclass-specificturned out to be cancer-related genes differentially regulated in all HCC patients, with quantitative ratherthan qualitative differences between the molecular subclasses. With the exception of a subset of samples with a definitive \beta-catenin gene signature, biological pathway analysis could not identify class-specific pathwaysreflecting the activation of distinct oncogenic programs. In conclusion, we have found that gene expressionprofiling of HCC biopsies has limited potential to direct therapies that target specific driver pathways, but canidentify subgroups of patients with different prognosis.

Authors

Zuzanna Makowska, Tujana Boldanova, David Adametz, Luca Quagliata, Julia E. Vogt, Michael T. Dill, Mathias S. Matter, Volker Roth, Luigi Terracciano, Markus H. Heim

Submitted

Journal of Pathology: Clinical Research, 2016

LinkDOI

Abstract

This paper proposes a new framework to find associations between somatic mu- tations and clinical features in cancer. The clinical features are directly extracted from the Electronic Health Records by performing a large-scale clustering of the sentences. Using a linear mixed model, we find significant associations between EHR-based phenotypes and gene mutations, while correcting for the cancer type as a confounding effect. To the author’s knowledge, this is the first attempt to per- form genetic association studies using EHR-based phenotypes. Such research has the potential to help in the discovery of unknown mechanisms in cancer, which will allow to prevent the disease, monitor patients at risk, and design tailored treatments for the patients.

Authors

Melanie F. Pradier, Stefan Stark, Stephanie Hyland, Julia E. Vogt, Gunnar Rätsch, and Fernando Perez-Cruz

Submitted

Paper + Spotlight Talk at Machine Learning for Computational Biology Workshop in Neural Information Processing Systems Conference 2015

Link

Authors

Melanie F. Pradier, Theofanis Karaletsos, Stefan Stark, Julia E. Vogt, Gunnar Rätsch, and Fernando Perez-Cruz

Submitted

Accepted Abstract at Machine Learning for Healthcare Workshop in Neural Information Processing Systems Conference 2015

Link

Abstract

We present a novel probabilistic clustering model for objects that are represented via pairwise distances and observed at different time points. The proposed method utilizes the information given by adjacent time points to find the underlying cluster structure and obtain a smooth cluster evolution. This approach allows the number of objects and clusters to differ at every time point, and no identification on the identities of the objects is needed. Further, the model does not require the number of clusters being specified in advance—they are instead determined automatically using a Dirichlet process prior. We validate our model on synthetic data showing that the proposed method is more accurate than state-of-the-art clustering methods. Finally, we use our dynamic clustering model to analyze and illustrate the evolution of brain cancer patients over time.

Authors

Julia E. Vogt, Marius Kloft, Stefan Stark, Sandhya Prabhakaran, Sudhir Raman, Volker Roth and Gunnar Rätsch

Submitted

Machine Learning Journal, 2015

LinkDOI

Abstract

A major challenge in computational biology is to find simple representations of high-dimensional data that best reveal the underlying structure. In this work, we present an intuitive and easy-to-implement method based on ranked neighborhood comparisons that detects structure in unsupervised data. The method is based on ordering objects in terms of similarity and on the mutual overlap of nearest neighbors. This basic framework was originally introduced in the field of social network analysis to detect actor communities. We demonstrate that the same ideas can successfully be applied to biomedical data sets in order to reveal complex underlying structure. The algorithm is very efficient and works on distance data directly without requiring a vectorial embedding of data. Comprehensive experiments demonstrate the validity of this approach. Comparisons with state-of-the-art clustering methods show that the presented method outperforms hierarchical methods as well as density based clustering methods and model-based clustering. A further advantage of the method is that it simultaneously provides a visualization of the data. Especially in biomedical applications, the visualization of data can be used as a first pre-processing step when analyzing real world data sets to get an intuition of the underlying data structure. We apply this model to synthetic data as well as to various biomedical data sets which demonstrate the high quality and usefulness of the inferred structure.

Authors

Julia E. Vogt

Submitted

IEEE/ACM Transactions on Computational Biology and Bioinformatics (Volume: 12 , Issue: 4 , July-Aug. 1 2015)

LinkDOI

Abstract

The use of pegylated interferon-\alpha (pegIFN-\alpha) has replaced unmodified recombinant IFN-\alpha for the treatment of chronic viral hepatitis. While the superior antiviral efficacy of pegIFN-\alpha is generally attributed to improved pharmacokinetic properties, the pharmacodynamic effects of pegIFN-\alpha in the liver have not been studied. Here, we analyzed pegIFN-\alpha–induced signaling and gene regulation in paired liver biopsies obtained prior to treatment and during the first week following pegIFN-\alpha injection in 18 patients with chronic hepatitis C. Despite sustained high concentrations of pegIFN-\alpha in serum, the Jak/STAT pathway was activated in hepatocytes only on the first day after pegIFN-\alpha administration. Evaluation of liver biopsies revealed that pegIFN-\alpha induces hundreds of genes that can be classified into four clusters based on different temporal expression profiles. In all clusters, gene transcription was mainly driven by IFN-stimulated gene factor 3 (ISGF3). Compared with conventional IFN-\alpha therapy, pegIFN-\alpha induced a broader spectrum of gene expression, including many genes involved in cellular immunity. IFN-induced secondary transcription factors did not result in additional waves of gene expression. Our data indicate that the superior antiviral efficacy of pegIFN-\alpha is not the result of prolonged Jak/STAT pathway activation in hepatocytes, but rather is due to induction of additional genes that are involved in cellular immune responses.

Authors

Michael T. Dill, Zuzanna Makowska, Gaia Trincucci, Andreas J. Gruber, Julia E. Vogt, Magdalena Filipowicz, Diego Calabrese, Ilona Krol, Daryl T. Lau, Luigi Terracciano, Erik van Nimwegen, Volker Roth and Markus H. Heim

Submitted

The Journal of Clinical Investigation

LinkDOI

Abstract

We present a Bayesian approach for estimating the relative frequencies of multi-single nucleotide polymorphism (SNP) haplotypes in populations of the malaria parasite Plasmodium falciparum by using microarray SNP data from human blood samples. Each sample comes from a malaria patient and contains one or several parasite clones that may genetically differ. Samples containing multiple parasite clones with different genetic markers pose a special challenge. The situation is comparable with a polyploid organism. The data from each blood sample indicates whether the parasites in the blood carry a mutant or a wildtype allele at various selected genomic positions. If both mutant and wildtype alleles are detected at a given position in a multiply infected sample, the data indicates the presence of both alleles, but the ratio is unknown. Thus, the data only partially reveals which specific combinations of genetic markers (i.e. haplotypes across the examined SNPs) occur in distinct parasite clones. In addition, SNP data may contain errors at non-negligible rates. We use a multinomial mixture model with partially missing observations to represent this data and a Markov chain Monte Carlo method to estimate the haplotype frequencies in a population. Our approach addresses both challenges, multiple infections and data errors.

Authors

Leonore Wigger, Julia E. Vogt, Volker Roth

Submitted

Statistics in Medicine: 04/2013

LinkDOI

Abstract

Partitioning methods for observations represented by pairwise dissimilarities are studied. Particular emphasis is put on their properties when applied to dissimilarity matrices that do not admit a loss-free embedding into a vector space. Specifically, the Pairwise Clustering cost function is shown to exhibit a shift invariance property which basically means that any symmetric dissimilarity matrix can be modified to allow a vector-space representation without distorting the optimal group structure. In an approximate sense, the same holds true for a probabilistic generalization of Pairwise Clustering, the so-called Wishart–Dirichlet Cluster Process. This shift-invariance property essentially means that these clustering methods are “blind” against Euclidean or metric violations. From the application side, such blindness against metric violations might be seen as a highly desired feature, since it broadens the applicability of certain algorithms. From the viewpoint of theory building, however, the same property might be viewed as a “negative” result, since studying these algorithms will not lead to any new insights on the role of metricity in clustering problems.

Authors

Volker Roth, Thomas J. Fuchs, Julia E. Vogt, Sandhya Prabhakaran, Joachim M. Buhmann

Submitted

Similarity-Based Pattern Analysis and Recognition, 157-177

LinkDOI

Abstract

Introduction IFN-\alpha signals through the Jak-STAT pathway to induce expression of IFN-stimulated genes (ISGs) with antiviral functions. USP18 is an IFN-inducible negative regulator of the Jak-STAT pathway. Upregulation of USP18 results in a long-lasting desensitization of IFN-\alpha signalling. As a result of this IFN-induced refractoriness, ISG levels decrease back to baseline despite continuous presence of the cytokine. Pegylated forms of IFN-\alpha (pegIFN-\alpha) are currently in clinical use for treatment of chronic hepatitis C virus infection. PegIFN-\alphas show increased anti-hepatitis C virus efficacy compared to nonpegylated IFN-\alpha. This has been attributed to the significantly longer plasma half-life of the pegylated form. However, the underlying assumption that persistently high plasma levels obtained with pegIFN-\alpha therapy result in ongoing stimulation of ISGs in the liver has never been tested. In the present study we therefore investigated the kinetics of Jak-STAT pathway activation and ISG induction in the human liver at several time points during the first week of pegIFN-\alpha therapy. Methods 18 patients with chronic hepatitis C underwent a liver biopsy 4 h (n = 6), 16 h, 48 h, 96 h or 144 h (all n = 3) after the first injection of pegIFN-\alpha-2b. Additional 3 patients received pegIFN-\alpha-2a and were biopsied at 144 h. The activation of Jak-STAT signalling and USP18 upregulation were assessed by immunohistochemistry and Western blot. Gene expression analysis was performed using Human Genome U133 Plus 2.0 arrays and Bioconductor packages of R statistical environment. Results A single dose of pegIFN-\alpha-2b resulted in elevated IFN-\alpha plasma levels throughout the one-week dosing interval. Despite the continuous IFN-\alpha exposure, strong activation of the Jak-STAT pathway was only observed at early time points after administration. Almost 500 genes were significantly upregulated in the liver samples following pegIFN-\alpha stimulation. The breadth of transcriptional response to pegIFN-\alpha was maximal 16 h post-injection and decreased gradually, with only few genes significantly upregulated after 144 h of treatment. Bayesian clustering of the gene expression data revealed 4 distinct groups of the ISGs based on the temporal patterns of regulation. Of 494 upregulated ISGs, the expression of 474 peaked 4 h or 16 h after pegIFN-\alpha administration, followed by a steady decline of mRNA levels through the remaining 128 h of treatment. This transient activation of the Jak-STAT pathway coincided with elevated expression of USP18 on the protein level, which was first detectable 16 post-injection. Conclusion PegIFN-\alpha induces a transient activation of Jak-STAT signalling and ISG upregulation in human liver, in spite of persistent high serum concentrations. The short-lived STAT1 phosphorylation and gene induction can be explained by upregulation of USP18 and establishment of refractory state. The superior efficacy of pegIFN-\alpha compared to conventional IFN-\alpha for chronic hepatitis C therapy cannot be explained by persistent signalling and ISG induction during the one-week dosing interval.

Authors

Z. Makowska, M. T. Dill, Julia E. Vogt, Magdalena Filipowicz Sinnreich, L. Terraciano, Volker Roth, M. H. Heim

Submitted

Cytokine 59(3):563–564, 2012

LinkDOI

Abstract

Archetype analysis involves the identification of representative objects from amongst a set of multivariate data such that the data can be expressed as a convex combination of these representative objects. Existing methods for archetype analysis assume a fixed number of archetypes a priori. Multiple runs of these methods for different choices of archetypes are required for model selection. Not only is this computationally infeasible for larger datasets, in heavy-noise settings model selection becomes cumbersome. In this paper, we present a novel extension to these existing methods with the specific focus of relaxing the need to provide a fixed number of archetypes beforehand. Our fast iterative optimization algorithm is devised to automatically select the right model using BIC scores and can easily be scaled to noisy, large datasets. These benefits are achieved by introducing a Group-Lasso component popular for sparse linear regression. The usefulness of the approach is demonstrated through simulations and on a real world application of document analysis for identifying topics.

Authors

Sandhya Prabhakaran, Sudhir Raman, Julia E. Vogt, Volker Roth

Submitted

Pattern Recognition: Joint 34th DAGM and 36th OAGM Symposium, Lecture Notes in Computer Science, 2012

LinkDOI

Abstract

The Group-Lasso is a well-known tool for joint regularization in machine learning methods. While the l_{1,2} and the l_{1,\infty} version have been studied in detail and efficient algorithms exist, there are still open questions regarding other l_{1,p} variants. We characterize conditions for solutions of the l_{1,p} Group-Lasso for all p-norms with 1 <= p <= \infty, and we present a unified active set algorithm. For all p-norms, a highly efficient projected gradient algorithm is presented. This new algorithm enables us to compare the prediction performance of many variants of the Group-Lasso in a multi-task learning setting, where the aim is to solve many learning problems in parallel which are coupled via the Group-Lasso constraint. We conduct large-scale experiments on synthetic data and on two real-world data sets. In accordance with theoretical characterizations of the different norms we observe that the weak-coupling norms with p between 1.5 and 2 consistently outperform the strong-coupling norms with p >> 2.

Authors

Julia E. Vogt, Volker Roth

Submitted

ICML 2012: Proceedings of the 29th international conference on Machine Learning

Link

Abstract

BACKGROUND & AIMS: The host immune response during the chronic phase of hepatitis C virus infection varies among individuals; some patients have a no interferon (IFN) response in the liver, whereas others have full activation of IFN-stimulated genes (ISGs). Preactivation of this endogenous IFN system is associated with nonresponse to pegylated IFN-\alpha (pegIFN-\alpha) and ribavirin. Genome-wide association studies have associated allelic variants near the IL28B (IFN\lambda3) gene with treatment response. We investigated whether IL28B genotype determines the constitutive expression of ISGs in the liver and compared the abilities of ISG levels and IL28B genotype to predict treatment outcome. METHODS: We genotyped 109 patients with chronic hepatitis C for IL28B allelic variants and quantified the hepatic expression of ISGs and of IL28B. Decision tree ensembles, in the form of a random forest classifier, were used to calculate the relative predictive power of these different variables in a multivariate analysis. RESULTS: The minor IL28B allele was significantly associated with increased expression of ISG. However, stratification of the patients according to treatment response revealed increased ISG expression in nonresponders, irrespective of IL28B genotype. Multivariate analysis of ISG expression, IL28B genotype, and several other factors associated with response to therapy identified ISG expression as the best predictor of treatment response. CONCLUSIONS: IL28B genotype and hepatic expression of ISGs are independent predictors of response to treatment with pegIFN-\alpha and ribavirin in patients with chronic hepatitis C. The most accurate prediction of response was obtained with a 4-gene classifier comprising IFI27, ISG15, RSAD2, and HTATIP2.

Authors

Michael T. Dill, Francois H.T. Duong, Julia E. Vogt, Stephanie Bibert, Pierre-Yves Bochud, Luigi Terracciano, Andreas Papassotiropoulos, Volker Roth and Markus H. Heim

Submitted

Gastroenterology, 2011 Mar;140(3):1021-1031.e10

LinkDOI

Abstract

The l_{1,\infty} norm and the l_{1,2} norm are well known tools for joint regularization in Group-Lasso methods. While the l_{1,2} version has been studied in detail, there are still open questions regarding the uniqueness of solutions and the efficiency of algorithms for the l_{1,\infty} variant. For the latter, we characterize the conditions for uniqueness of solutions, we present a simple test for uniqueness, and we derive a highly efficient active set algorithm that can deal with input dimensions in the millions. We compare both variants of the Group-Lasso for the two most common application scenarios of the Group-Lasso, one is to obtain sparsity on the level of groups in “standard” prediction problems, the second one is multi-task learning where the aim is to solve many learning problems in parallel which are coupled via the Group-Lasso constraint. We show that both version perform quite similar in “standard” applications. However, a very clear distinction between the variants occurs in multi-task settings where the l_{1,2} version consistently outperforms the l_{1,\infty} counterpart in terms of prediction accuracy.

Authors

Julia E. Vogt, Volker Roth

Submitted

Pattern Recognition: 32-nd DAGM Symposium, Lecture Notes in Computer Science, 2010

Link

Abstract

We present a probabilistic model for clustering of objects represented via pairwise dissimilarities. We propose that even if an underlying vectorial representation exists, it is better to work directly with the dissimilarity matrix hence avoiding unnecessary bias and variance caused by embeddings. By using a Dirichlet process prior we are not obliged to fix the number of clusters in advance. Furthermore, our clustering model is permutation-, scale- and translation-invariant, and it is called the Translation-invariant Wishart Dirichlet (TIWD) process. A highly efficient MCMC sampling algorithm is presented. Experiments show that the TIWD process exhibits several advantages over competing approaches.

Authors

Julia E. Vogt, Sandhya Prabhakaran, Thomas J. Fuchs, Volker Roth

Submitted

ICML 2010: Proceedings of the 27th international conference on Machine Learning

Link