Shyam Visweswaran

Yanwu Xu, Li Sun, Wei Peng et al.

This paper introduces an innovative methodology for producing high-quality 3D lung CT images guided by textual information. While diffusion-based generative models are increasingly used in medical imaging, current state-of-the-art approaches are limited to low-resolution outputs and underutilize radiology reports' abundant information. The radiology reports can enhance the generation process by providing additional guidance and offering fine-grained control over the synthesis of images. Nevertheless, expanding text-guided generation to high-resolution 3D images poses significant memory and anatomical detail-preserving challenges. Addressing the memory issue, we introduce a hierarchical scheme that uses a modified UNet architecture. We start by synthesizing low-resolution images conditioned on the text, serving as a foundation for subsequent generators for complete volumetric data. To ensure the anatomical plausibility of the generated samples, we provide further guidance by generating vascular, airway, and lobular segmentation masks in conjunction with the CT images. The model demonstrates the capability to use textual input and segmentation tasks to generate synthesized images. The results of comparative assessments indicate that our approach exhibits superior performance compared to the most advanced models based on GAN and diffusion techniques, especially in accurately retaining crucial anatomical features such as fissure lines, airways, and vascular structures. This innovation introduces novel possibilities. This study focuses on two main objectives: (1) the development of a method for creating images based on textual prompts and anatomical components, and (2) the capability to generate new images conditioning on anatomical elements. The advancements in image generation can be applied to enhance numerous downstream tasks.

Sonish Sivarajkumar, Mark Kelley, Alyssa Samolyk-Mazzanti et al.

Large language models (LLMs) have shown remarkable capabilities in Natural Language Processing (NLP), especially in domains where labeled data is scarce or expensive, such as clinical domain. However, to unlock the clinical knowledge hidden in these LLMs, we need to design effective prompts that can guide them to perform specific clinical NLP tasks without any task-specific training data. This is known as in-context learning, which is an art and science that requires understanding the strengths and weaknesses of different LLMs and prompt engineering approaches. In this paper, we present a comprehensive and systematic experimental study on prompt engineering for five clinical NLP tasks: Clinical Sense Disambiguation, Biomedical Evidence Extraction, Coreference Resolution, Medication Status Extraction, and Medication Attribute Extraction. We assessed the prompts proposed in recent literature, including simple prefix, simple cloze, chain of thought, and anticipatory prompts, and introduced two new types of prompts, namely heuristic prompting and ensemble prompting. We evaluated the performance of these prompts on three state-of-the-art LLMs: GPT-3.5, BARD, and LLAMA2. We also contrasted zero-shot prompting with few-shot prompting, and provide novel insights and guidelines for prompt engineering for LLMs in clinical NLP. To the best of our knowledge, this is one of the first works on the empirical evaluation of different prompt engineering approaches for clinical NLP in this era of generative AI, and we hope that it will inspire and inform future research in this area.

Sonish Sivarajkumar, Fengyi Gao, Parker E. Denny et al.

Post-stroke patient rehabilitation requires precise, personalized treatment plans. Natural Language Processing (NLP) offers potential to extract valuable exercise information from clinical notes, aiding in the development of more effective rehabilitation strategies. Objective: This study aims to develop and evaluate a variety of NLP algorithms to extract and categorize physical rehabilitation exercise information from the clinical notes of post-stroke patients treated at the University of Pittsburgh Medical Center. A cohort of 13,605 patients diagnosed with stroke was identified, and their clinical notes containing rehabilitation therapy notes were retrieved. A comprehensive clinical ontology was created to represent various aspects of physical rehabilitation exercises. State-of-the-art NLP algorithms were then developed and compared, including rule-based, machine learning-based algorithms, and large language model (LLM)-based algorithms (ChatGPT). Analysis was conducted on a dataset comprising 23,724 notes with detailed demographic and clinical characteristics. The rule-based NLP algorithm demonstrated superior performance in most areas, particularly in detecting the 'Right Side' location with an F1 score of 0.975, outperforming Gradient Boosting by 0.063. Gradient Boosting excelled in 'Lower Extremity' location detection (F1 score: 0.978), surpassing rule-based NLP by 0.023. It also showed notable performance in 'Passive Range of Motion' with an F1 score of 0.970, a 0.032 improvement over rule-based NLP. The rule-based algorithm efficiently handled 'Duration', 'Sets', and 'Reps' with F1 scores up to 0.65. LLM-based NLP, particularly ChatGPT with few-shot prompts, achieved high recall but generally lower precision and F1 scores. However, it notably excelled in 'Backward Plane' motion detection, achieving an F1 score of 0.846, surpassing the rule-based algorithm's 0.720.

Milos Hauskrecht, Michal Valko, Shyam Visweswaran et al.

We develop and evaluate a data-driven approach for detecting unusual (anomalous) patient-management actions using past patient cases stored in an electronic health record (EHR) system. Our hypothesis is that patient-management actions that are unusual with respect to past patients may be due to a potential error and that it is worthwhile to raise an alert if such a condition is encountered. We evaluate this hypothesis using data obtained from the electronic health records of 4,486 post-cardiac surgical patients. We base the evaluation on the opinions of a panel of experts. The results support that anomaly-based alerting can have reasonably low false alert rates and that stronger anomalies are correlated with higher alert rates.

Ke Yu, Shyam Visweswaran, Kayhan Batmanghelich

Learning accurate drug representations is essential for task such as computational drug repositioning. A drug hierarchy is a valuable source that encodes knowledge of relations among drugs in a tree-like structure where drugs that act on the same organs, treat the same disease, or bind to the same biological target are grouped together. However, its utility in learning drug representations has not yet been explored, and currently described drug representations cannot place novel molecules in a drug hierarchy. Here, we develop a semi-supervised drug embedding that incorporates two sources of information: (1) underlying chemical grammar that is inferred from chemical structures of drugs and drug-like molecules (unsupervised), and (2) hierarchical relations that are encoded in an expert-crafted hierarchy of approved drugs (supervised). We use the Variational Auto-Encoder (VAE) framework to encode the chemical structures of molecules and use the drug-drug similarity information obtained from the hierarchy to induce the clustering of drugs in hyperbolic space. The hyperbolic space is amenable for encoding hierarchical relations. Our qualitative results support that the learned drug embedding can induce the hierarchical relations among drugs. We demonstrate that the learned drug embedding can be used for drug repositioning.

Milos Hauskrecht, Michal Valko, Branislav Kveton et al.

Anomaly detection methods can be very useful in identifying interesting or concerning events. In this work, we develop and examine new probabilistic anomaly detection methods that let us evaluate management decisions for a specific patient and identify those decisions that are highly unusual with respect to patients with the same or similar condition. The statistics used in this detection are derived from probabilistic models such as Bayesian networks that are learned from a database of past patient cases. We apply our methods to the problem of identifying unusual patient-management decisions in post-surgical cardiac patients.

Michal Valko, Gregory Cooper, Amy Seybert et al.

Anomaly detection methods can be very useful in identifying unusual or interesting patterns in data. A recently proposed conditional anomaly detection framework extends anomaly detection to the problem of identifying anomalous patterns on a subset of attributes in the data. The anomaly always depends (is conditioned) on the value of remaining attributes. The work presented in this paper focuses on instance-based methods for detecting conditional anomalies. The methods rely on the distance metric to identify examples in the dataset that are most critical for detecting the anomaly. We investigate various metrics and metric learning methods to optimize the performance of the instance-based anomaly detection methods. We show the benefits of the instance-based methods on two real-world detection problems: detection of unusual admission decisions for patients with the community-acquired pneumonia and detection of unusual orders of an HPF4 test that is used to confirm Heparin induced thrombocytopenia - a life-threatening condition caused by the Heparin therapy.

Xizhi Wu, Nelly Estefanie Garduno-Rapp, Justin F Rousseau et al.

Unlike most primary headaches, secondary headaches need specialized care and can have devastating consequences if not treated promptly. Clinical guidelines highlight several 'red flag' features, such as thunderclap onset, meningismus, papilledema, focal neurologic deficits, signs of temporal arteritis, systemic illness, and the 'worst headache of their life' presentation. Despite these guidelines, determining which patients require urgent evaluation remains challenging in primary care settings. Clinicians often work with limited time, incomplete information, and diverse symptom presentations, which can lead to under-recognition and inappropriate care. We present a large language model (LLM)-based multi-agent clinical decision support system built on an orchestrator-specialist architecture, designed to perform explicit and interpretable secondary headache diagnosis from free-text clinical vignettes. The multi-agent system decomposes diagnosis into seven domain-specialized agents, each producing a structured and evidence-grounded rationale, while a central orchestrator performs task decomposition and coordinates agent routing. We evaluated the multi-agent system using 90 expert-validated secondary headache cases and compared its performance with a single-LLM baseline across two prompting strategies: question-based prompting (QPrompt) and clinical practice guideline-based prompting (GPrompt). We tested five open-source LLMs (Qwen-30B, GPT-OSS-20B, Qwen-14B, Qwen-8B, and Llama-3.1-8B), and found that the orchestrated multi-agent system with GPrompt consistently achieved the highest F1 scores, with larger gains in smaller models. These findings demonstrate that structured multi-agent reasoning improves accuracy beyond prompt engineering alone and offers a transparent, clinically aligned approach for explainable decision support in secondary headache diagnosis.

Shantanu Ghosh, Clare B. Poynton, Shyam Visweswaran et al. · amazon-science, cmu

The lack of large and diverse training data on Computer-Aided Diagnosis (CAD) in breast cancer detection has been one of the concerns that impedes the adoption of the system. Recently, pre-training with large-scale image text datasets via Vision-Language models (VLM) (\eg CLIP) partially addresses the issue of robustness and data efficiency in computer vision (CV). This paper proposes Mammo-CLIP, the first VLM pre-trained on a substantial amount of screening mammogram-report pairs, addressing the challenges of dataset diversity and size. Our experiments on two public datasets demonstrate strong performance in classifying and localizing various mammographic attributes crucial for breast cancer detection, showcasing data efficiency and robustness similar to CLIP in CV. We also propose Mammo-FActOR, a novel feature attribution method, to provide spatial interpretation of representation with sentence-level granularity within mammography reports. Code is available publicly: \url{https://github.com/batmanlab/Mammo-CLIP}.

Sonish Sivarajkumar, Thomas Yu CHow Tam, Haneef Ahamed Mohammad et al.

Alzheimer's Disease (AD) is the most common form of dementia in the United States. Sleep is one of the lifestyle-related factors that has been shown critical for optimal cognitive function in old age. However, there is a lack of research studying the association between sleep and AD incidence. A major bottleneck for conducting such research is that the traditional way to acquire sleep information is time-consuming, inefficient, non-scalable, and limited to patients' subjective experience. A gold standard dataset is created from manual annotation of 570 randomly sampled clinical note documents from the adSLEEP, a corpus of 192,000 de-identified clinical notes of 7,266 AD patients retrieved from the University of Pittsburgh Medical Center (UPMC). We developed a rule-based Natural Language Processing (NLP) algorithm, machine learning models, and Large Language Model(LLM)-based NLP algorithms to automate the extraction of sleep-related concepts, including snoring, napping, sleep problem, bad sleep quality, daytime sleepiness, night wakings, and sleep duration, from the gold standard dataset. Rule-based NLP algorithm achieved the best performance of F1 across all sleep-related concepts. In terms of Positive Predictive Value (PPV), rule-based NLP algorithm achieved 1.00 for daytime sleepiness and sleep duration, machine learning models: 0.95 and for napping, 0.86 for bad sleep quality and 0.90 for snoring; and LLAMA2 with finetuning achieved PPV of 0.93 for Night Wakings, 0.89 for sleep problem, and 1.00 for sleep duration. The results show that the rule-based NLP algorithm consistently achieved the best performance for all sleep concepts. This study focused on the clinical notes of patients with AD, but could be extended to general sleep information extraction for other diseases.

Thomas Yu Chow Tam, Sonish Sivarajkumar, Sumit Kapoor et al.

With generative artificial intelligence (AI), particularly large language models (LLMs), continuing to make inroads in healthcare, it is critical to supplement traditional automated evaluations with human evaluations. Understanding and evaluating the output of LLMs is essential to assuring safety, reliability, and effectiveness. However, human evaluation's cumbersome, time-consuming, and non-standardized nature presents significant obstacles to comprehensive evaluation and widespread adoption of LLMs in practice. This study reviews existing literature on human evaluation methodologies for LLMs in healthcare. We highlight a notable need for a standardized and consistent human evaluation approach. Our extensive literature search, adhering to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, includes publications from January 2018 to February 2024. The review examines the human evaluation of LLMs across various medical specialties, addressing factors such as evaluation dimensions, sample types and sizes, selection, and recruitment of evaluators, frameworks and metrics, evaluation process, and statistical analysis type. Drawing on the diverse evaluation strategies employed in these studies, we propose a comprehensive and practical framework for human evaluation of LLMs: QUEST: Quality of Information, Understanding and Reasoning, Expression Style and Persona, Safety and Harm, and Trust and Confidence. This framework aims to improve the reliability, generalizability, and applicability of human evaluation of LLMs in different healthcare applications by defining clear evaluation dimensions and offering detailed guidelines.

Eric V. Strobl, Shyam Visweswaran

Personalized medicine seeks to identify the causal effect of treatment for a particular patient as opposed to a clinical population at large. Most investigators estimate such personalized treatment effects by regressing the outcome of a randomized clinical trial (RCT) on patient covariates. The realized value of the outcome may however lie far from the conditional expectation. We therefore introduce a method called Dirac Delta Regression (DDR) that estimates the entire conditional density from RCT data in order to visualize the probabilities across all possible outcome values. DDR transforms the outcome into a set of asymptotically Dirac delta distributions and then estimates the density using non-linear regression. The algorithm can identify significant differences in patient-specific outcomes even when no population level effect exists. Moreover, DDR outperforms state-of-the-art algorithms in conditional density estimation by a large margin even in the small sample regime. An R package is available at https://github.com/ericstrobl/DDR.

Eric V. Strobl, Shyam Visweswaran

Motivation: Algorithms that discover variables which are causally related to a target may inform the design of experiments. With observational gene expression data, many methods discover causal variables by measuring each variable's degree of statistical dependence with the target using dependence measures (DMs). However, other methods measure each variable's ability to explain the statistical dependence between the target and the remaining variables in the data using conditional dependence measures (CDMs), since this strategy is guaranteed to find the target's direct causes, direct effects, and direct causes of the direct effects in the infinite sample limit. In this paper, we design a new algorithm in order to systematically compare the relative abilities of DMs and CDMs in discovering causal variables from gene expression data. Results: The proposed algorithm using a CDM is sample efficient, since it consistently outperforms other state-of-the-art local causal discovery algorithms when samples sizes are small. However, the proposed algorithm using a CDM outperforms the proposed algorithm using a DM only when sample sizes are above several hundred. These results suggest that accurate causal discovery from gene expression data using current CDM-based algorithms requires datasets with at least several hundred samples. Availability: The proposed algorithm is freely available at https://github.com/ericstrobl/DvCD.

Shantanu Ghosh, Vedant Parthesh Joshi, Rayan Syed et al.

Breast cancer is one of the leading causes of death among women worldwide. We introduce Mammo-FM, the first foundation model specifically for mammography, pretrained on the largest and most diverse dataset to date - 140,677 patients (821,326 mammograms) across four U.S. institutions. Mammo-FM provides a unified foundation for core clinical tasks in breast imaging, including cancer diagnosis, pathology localization, structured report generation, and cancer risk prognosis within a single framework. Its alignment between images and text enables both visual and textual interpretability, improving transparency and clinical auditability, which are essential for real-world adoption. We rigorously evaluate Mammo-FM across diagnosis, prognosis, and report-generation tasks in in- and out-of-distribution datasets. Despite operating on native-resolution mammograms and using only one-third of the parameters of state-of-the-art generalist FMs, Mammo-FM consistently outperforms them across multiple public and private benchmarks. These results highlight the efficiency and value of domain-specific foundation models designed around the full spectrum of tasks within a clinical domain and emphasize the importance of rigorous, domain-aligned evaluation.

Fengyi Gao, Xingyu Zhang, Sonish Sivarajkumar et al.

In this study, we utilized statistical analysis and machine learning methods to examine whether rehabilitation exercises can improve patients post-stroke functional abilities, as well as forecast the improvement in functional abilities. Our dataset is patients' rehabilitation exercises and demographic information recorded in the unstructured electronic health records (EHRs) data and free-text rehabilitation procedure notes. We collected data for 265 stroke patients from the University of Pittsburgh Medical Center. We employed a pre-existing natural language processing (NLP) algorithm to extract data on rehabilitation exercises and developed a rule-based NLP algorithm to extract Activity Measure for Post-Acute Care (AM-PAC) scores, covering basic mobility (BM) and applied cognitive (AC) domains, from procedure notes. Changes in AM-PAC scores were classified based on the minimal clinically important difference (MCID), and significance was assessed using Friedman and Wilcoxon tests. To identify impactful exercises, we used Chi-square tests, Fisher's exact tests, and logistic regression for odds ratios. Additionally, we developed five machine learning models-logistic regression (LR), Adaboost (ADB), support vector machine (SVM), gradient boosting (GB), and random forest (RF)-to predict outcomes in functional ability. Statistical analyses revealed significant associations between functional improvements and specific exercises. The RF model achieved the best performance in predicting functional outcomes. In this study, we identified three rehabilitation exercises that significantly contributed to patient post-stroke functional ability improvement in the first two months. Additionally, the successful application of a machine learning model to predict patient-specific functional outcomes underscores the potential for precision rehabilitation.

David Oniani, Xizhi Wu, Shyam Visweswaran et al.

Background Large Language Models (LLMs), enhanced with Clinical Practice Guidelines (CPGs), can significantly improve Clinical Decision Support (CDS). However, methods for incorporating CPGs into LLMs are not well studied. Methods We develop three distinct methods for incorporating CPGs into LLMs: Binary Decision Tree (BDT), Program-Aided Graph Construction (PAGC), and Chain-of-Thought-Few-Shot Prompting (CoT-FSP). To evaluate the effectiveness of the proposed methods, we create a set of synthetic patient descriptions and conduct both automatic and human evaluation of the responses generated by four LLMs: GPT-4, GPT-3.5 Turbo, LLaMA, and PaLM 2. Zero-Shot Prompting (ZSP) was used as the baseline method. We focus on CDS for COVID-19 outpatient treatment as the case study. Results All four LLMs exhibit improved performance when enhanced with CPGs compared to the baseline ZSP. BDT outperformed both CoT-FSP and PAGC in automatic evaluation. All of the proposed methods demonstrated high performance in human evaluation. Conclusion LLMs enhanced with CPGs demonstrate superior performance, as compared to plain LLMs with ZSP, in providing accurate recommendations for COVID-19 outpatient treatment, which also highlights the potential for broader applications beyond the case study.

Ke Yu, Shyam Visweswaran, Kayhan Batmanghelich

Learning accurate drug representation is essential for tasks such as computational drug repositioning and prediction of drug side-effects. A drug hierarchy is a valuable source that encodes human knowledge of drug relations in a tree-like structure where drugs that act on the same organs, treat the same disease, or bind to the same biological target are grouped together. However, its utility in learning drug representations has not yet been explored, and currently described drug representations cannot place novel molecules in a drug hierarchy. Here, we develop a semi-supervised drug embedding that incorporates two sources of information: (1) underlying chemical grammar that is inferred from molecular structures of drugs and drug-like molecules (unsupervised), and (2) hierarchical relations that are encoded in an expert-crafted hierarchy of approved drugs (supervised). We use the Variational Auto-Encoder (VAE) framework to encode the chemical structures of molecules and use the knowledge-based drug-drug similarity to induce the clustering of drugs in hyperbolic space. The hyperbolic space is amenable for encoding hierarchical concepts. Both quantitative and qualitative results support that the learned drug embedding can accurately reproduce the chemical structure and induce the hierarchical relations among drugs. Furthermore, our approach can infer the pharmacological properties of novel molecules by retrieving similar drugs from the embedding space. We demonstrate that the learned drug embedding can be used to find new uses for existing drugs and to discover side-effects. We show that it significantly outperforms baselines in both tasks.

Eric V. Strobl, Shyam Visweswaran, Peter L. Spirtes

Many real datasets contain values missing not at random (MNAR). In this scenario, investigators often perform list-wise deletion, or delete samples with any missing values, before applying causal discovery algorithms. List-wise deletion is a sound and general strategy when paired with algorithms such as FCI and RFCI, but the deletion procedure also eliminates otherwise good samples that contain only a few missing values. In this report, we show that we can more efficiently utilize the observed values with test-wise deletion while still maintaining algorithmic soundness. Here, test-wise deletion refers to the process of list-wise deleting samples only among the variables required for each conditional independence (CI) test used in constraint-based searches. Test-wise deletion therefore often saves more samples than list-wise deletion for each CI test, especially when we have a sparse underlying graph. Our theoretical results show that test-wise deletion is sound under the justifiable assumption that none of the missingness mechanisms causally affect each other in the underlying causal graph. We also find that FCI and RFCI with test-wise deletion outperform their list-wise deletion and imputation counterparts on average when MNAR holds in both synthetic and real data.

Eric V. Strobl, Kun Zhang, Shyam Visweswaran

Constraint-based causal discovery (CCD) algorithms require fast and accurate conditional independence (CI) testing. The Kernel Conditional Independence Test (KCIT) is currently one of the most popular CI tests in the non-parametric setting, but many investigators cannot use KCIT with large datasets because the test scales cubicly with sample size. We therefore devise two relaxations called the Randomized Conditional Independence Test (RCIT) and the Randomized conditional Correlation Test (RCoT) which both approximate KCIT by utilizing random Fourier features. In practice, both of the proposed tests scale linearly with sample size and return accurate p-values much faster than KCIT in the large sample size context. CCD algorithms run with RCIT or RCoT also return graphs at least as accurate as the same algorithms run with KCIT but with large reductions in run time.

Eric V. Strobl, Peter L. Spirtes, Shyam Visweswaran

The PC algorithm allows investigators to estimate a complete partially directed acyclic graph (CPDAG) from a finite dataset, but few groups have investigated strategies for estimating and controlling the false discovery rate (FDR) of the edges in the CPDAG. In this paper, we introduce PC with p-values (PC-p), a fast algorithm which robustly computes edge-specific p-values and then estimates and controls the FDR across the edges. PC-p specifically uses the p-values returned by many conditional independence tests to upper bound the p-values of more complex edge-specific hypothesis tests. The algorithm then estimates and controls the FDR using the bounded p-values and the Benjamini-Yekutieli FDR procedure. Modifications to the original PC algorithm also help PC-p accurately compute the upper bounds despite non-zero Type II error rates. Experiments show that PC-p yields more accurate FDR estimation and control across the edges in a variety of CPDAGs compared to alternative methods.

Eric V. Strobl, Shyam Visweswaran

Ridge regularized linear models (RRLMs), such as ridge regression and the SVM, are a popular group of methods that are used in conjunction with coefficient hypothesis testing to discover explanatory variables with a significant multivariate association to a response. However, many investigators are reluctant to draw causal interpretations of the selected variables due to the incomplete knowledge of the capabilities of RRLMs in causal inference. Under reasonable assumptions, we show that a modified form of RRLMs can get very close to identifying a subset of the Markov boundary by providing a worst-case bound on the space of possible solutions. The results hold for any convex loss, even when the underlying functional relationship is nonlinear, and the solution is not unique. Our approach combines ideas in Markov boundary and sufficient dimension reduction theory. Experimental results show that the modified RRLMs are competitive against state-of-the-art algorithms in discovering part of the Markov boundary from gene expression data.

Shyam Visweswaran, Gregory F. Cooper

Learning Markov blanket (MB) structures has proven useful in performing feature selection, learning Bayesian networks (BNs), and discovering causal relationships. We present a formula for efficiently determining the number of MB structures given a target variable and a set of other variables. As expected, the number of MB structures grows exponentially. However, we show quantitatively that there are many fewer MB structures that contain the target variable than there are BN structures that contain it. In particular, the ratio of BN structures to MB structures appears to increase exponentially in the number of variables.

Eric V. Strobl, Shyam Visweswaran

Developing feature selection algorithms that move beyond a pure correlational to a more causal analysis of observational data is an important problem in the sciences. Several algorithms attempt to do so by discovering the Markov blanket of a target, but they all contain a forward selection step which variables must pass in order to be included in the conditioning set. As a result, these algorithms may not consider all possible conditional multivariate combinations. We improve on this limitation by proposing a backward elimination method that uses a kernel-based conditional dependence measure to identify the Markov blanket in a fully multivariate fashion. The algorithm is easy to implement and compares favorably to other methods on synthetic and real datasets.

Eric Strobl, Shyam Visweswaran

Deep learning methods have predominantly been applied to large artificial neural networks. Despite their state-of-the-art performance, these large networks typically do not generalize well to datasets with limited sample sizes. In this paper, we take a different approach by learning multiple layers of kernels. We combine kernels at each layer and then optimize over an estimate of the support vector machine leave-one-out error rather than the dual objective function. Our experiments on a variety of datasets show that each layer successively increases performance with only a few base kernels.