Jenna Wiens

Trenton Chang, Michael W. Sjoding, Jenna Wiens

As machine learning (ML) models gain traction in clinical applications, understanding the impact of clinician and societal biases on ML models is increasingly important. While biases can arise in the labels used for model training, the many sources from which these biases arise are not yet well-studied. In this paper, we highlight disparate censorship (i.e., differences in testing rates across patient groups) as a source of label bias that clinical ML models may amplify, potentially causing harm. Many patient risk-stratification models are trained using the results of clinician-ordered diagnostic and laboratory tests of labels. Patients without test results are often assigned a negative label, which assumes that untested patients do not experience the outcome. Since orders are affected by clinical and resource considerations, testing may not be uniform in patient populations, giving rise to disparate censorship. Disparate censorship in patients of equivalent risk leads to undertesting in certain groups, and in turn, more biased labels for such groups. Using such biased labels in standard ML pipelines could contribute to gaps in model performance across patient groups. Here, we theoretically and empirically characterize conditions in which disparate censorship or undertesting affect model performance across subgroups. Our findings call attention to disparate censorship as a source of label bias in clinical ML models.

Aaman Rebello, Shengpu Tang, Jenna Wiens et al.

Off-policy evaluation (OPE) aims to estimate the benefit of following a counterfactual sequence of actions, given data collected from executed sequences. However, existing OPE estimators often exhibit high bias and high variance in problems involving large, combinatorial action spaces. We investigate how to mitigate this issue using factored action spaces i.e. expressing each action as a combination of independent sub-actions from smaller action spaces. This approach facilitates a finer-grained analysis of how actions differ in their effects. In this work, we propose a new family of "decomposed" importance sampling (IS) estimators based on factored action spaces. Given certain assumptions on the underlying problem structure, we prove that the decomposed IS estimators have less variance than their original non-decomposed versions, while preserving the property of zero bias. Through simulations, we empirically verify our theoretical results, probing the validity of various assumptions. Provided with a technique that can derive the action space factorisation for a given problem, our work shows that OPE can be improved "for free" by utilising this inherent problem structure.

Harry Rubin-Falcone, Joyce Lee, Jenna Wiens

In time-series forecasting, future target values may be affected by both intrinsic and extrinsic effects. When forecasting blood glucose, for example, intrinsic effects can be inferred from the history of the target signal alone (\textit{i.e.} blood glucose), but accurately modeling the impact of extrinsic effects requires auxiliary signals, like the amount of carbohydrates ingested. Standard forecasting techniques often assume that extrinsic and intrinsic effects vary at similar rates. However, when auxiliary signals are generated at a much lower frequency than the target variable (e.g., blood glucose measurements are made every 5 minutes, while meals occur once every few hours), even well-known extrinsic effects (e.g., carbohydrates increase blood glucose) may prove difficult to learn. To better utilize these \textit{sparse but informative variables} (SIVs), we introduce a novel encoder/decoder forecasting approach that accurately learns the per-timepoint effect of the SIV, by (i) isolating it from intrinsic effects and (ii) restricting its learned effect based on domain knowledge. On a simulated dataset pertaining to the task of blood glucose forecasting, when the SIV is accurately recorded our approach outperforms baseline approaches in terms of rMSE (13.07 [95% CI: 11.77,14.16] vs. 14.14 [12.69,15.27]). In the presence of a corrupted SIV, the proposed approach can still result in lower error compared to the baseline but the advantage is reduced as noise increases. By isolating their effects and incorporating domain knowledge, our approach makes it possible to better utilize SIVs in forecasting.

Donna Tjandra, Jenna Wiens

Noisy training labels can hurt model performance. Most approaches that aim to address label noise assume label noise is independent from the input features. In practice, however, label noise is often feature or \textit{instance-dependent}, and therefore biased (i.e., some instances are more likely to be mislabeled than others). E.g., in clinical care, female patients are more likely to be under-diagnosed for cardiovascular disease compared to male patients. Approaches that ignore this dependence can produce models with poor discriminative performance, and in many healthcare settings, can exacerbate issues around health disparities. In light of these limitations, we propose a two-stage approach to learn in the presence instance-dependent label noise. Our approach utilizes \textit{\anchor points}, a small subset of data for which we know the observed and ground truth labels. On several tasks, our approach leads to consistent improvements over the state-of-the-art in discriminative performance (AUROC) while mitigating bias (area under the equalized odds curve, AUEOC). For example, when predicting acute respiratory failure onset on the MIMIC-III dataset, our approach achieves a harmonic mean (AUROC and AUEOC) of 0.84 (SD [standard deviation] 0.01) while that of the next best baseline is 0.81 (SD 0.01). Overall, our approach improves accuracy while mitigating potential bias compared to existing approaches in the presence of instance-dependent label noise.

Erkin Ötleş, Brian T. Denton, Jenna Wiens

As data shift or new data become available, updating clinical machine learning models may be necessary to maintain or improve performance over time. However, updating a model can introduce compatibility issues when the behavior of the updated model does not align with user expectations, resulting in poor user-model team performance. Existing compatibility measures depend on model decision thresholds, limiting their applicability in settings where models are used to generate rankings based on estimated risk. To address this limitation, we propose a novel rank-based compatibility measure, $C^R$, and a new loss function that aims to optimize discriminative performance while encouraging good compatibility. Applied to a case study in mortality risk stratification leveraging data from MIMIC, our approach yields more compatible models while maintaining discriminative performance compared to existing model selection techniques, with an increase in $C^R$ of $0.019$ ($95\%$ confidence interval: $0.005$, $0.035$). This work provides new tools to analyze and update risk stratification models used in clinical care.

Shengpu Tang, Jenna Wiens

In applying reinforcement learning (RL) to high-stakes domains, quantitative and qualitative evaluation using observational data can help practitioners understand the generalization performance of new policies. However, this type of off-policy evaluation (OPE) is inherently limited since offline data may not reflect the distribution shifts resulting from the application of new policies. On the other hand, online evaluation by collecting rollouts according to the new policy is often infeasible, as deploying new policies in these domains can be unsafe. In this work, we propose a semi-offline evaluation framework as an intermediate step between offline and online evaluation, where human users provide annotations of unobserved counterfactual trajectories. While tempting to simply augment existing data with such annotations, we show that this naive approach can lead to biased results. Instead, we design a new family of OPE estimators based on importance sampling (IS) and a novel weighting scheme that incorporate counterfactual annotations without introducing additional bias. We analyze the theoretical properties of our approach, showing its potential to reduce both bias and variance compared to standard IS estimators. Our analyses reveal important practical considerations for handling biased, noisy, or missing annotations. In a series of proof-of-concept experiments involving bandits and a healthcare-inspired simulator, we demonstrate that our approach outperforms purely offline IS estimators and is robust to imperfect annotations. Our framework, combined with principled human-centered design of annotation solicitation, can enable the application of RL in high-stakes domains.

Donna Tjandra, Trenton Chang, Sonali Parbhoo et al.

Objective: The growing availability of large-scale observational clinical datasets and challenges in conducting randomized controlled trials have spurred enthusiasm in using causal machine learning (ML) for causal inference in observational data. We present a roadmap for applying causal ML to observational data. Materials and methods: We outline the importance of assessing validity assumptions within available data and applying causal ML responsibly for clinical experts using causal ML and ML practitioners with limited clinical expertise. Observations: Despite advances in causal ML, its limitations remain largely under-appreciated across disciplines. This gap in shared knowledge may impact the validity of findings. Discussion: Causal assumptions must be satisfied and modeling choices justified. Otherwise, these approaches risk producing biased or misleading results, with consequences for clinical research and patient care. Conclusion: Causal ML can be a powerful tool for generating causal hypotheses. We provide a template to strengthen the rigor and interpretability of causal analyses.

Sarah Jabbour, Gregory Kondas, Ella Kazerooni et al.

We propose a permutation-based explanation method for image classifiers. Current image-model explanations like activation maps are limited to instance-based explanations in the pixel space, making it difficult to understand global model behavior. In contrast, permutation based explanations for tabular data classifiers measure feature importance by comparing model performance on data before and after permuting a feature. We propose an explanation method for image-based models that permutes interpretable concepts across dataset images. Given a dataset of images labeled with specific concepts like captions, we permute a concept across examples in the text space and then generate images via a text-conditioned diffusion model. Feature importance is then reflected by the change in model performance relative to unpermuted data. When applied to a set of concepts, the method generates a ranking of feature importance. We show this approach recovers underlying model feature importance on synthetic and real-world image classification tasks.

Trenton Chang, Tobias Schnabel, Adith Swaminathan et al.

Despite advances in large language models (LLMs) on reasoning and instruction-following benchmarks, it remains unclear whether they can reliably produce outputs aligned with a broad variety of user goals, a concept we refer to as steerability. The abundance of methods proposed to modify LLM behavior makes it unclear whether current LLMs are already steerable, or require further intervention. In particular, LLMs may exhibit (i) poor coverage, where rare user goals are underrepresented; (ii) miscalibration, where models overshoot requests; and (iii) side effects, where changes to one dimension of text inadvertently affect others. To systematically evaluate these failures, we introduce a framework based on a multi-dimensional goal space that models user goals and LLM outputs as vectors with dimensions corresponding to text attributes (e.g., reading difficulty). Applied to a text-rewriting task, we find that current LLMs struggle with steerability, as side effects are persistent. Interventions to improve steerability, such as prompt engineering, best-of-$N$ sampling, and reinforcement learning fine-tuning, have varying effectiveness, yet side effects remain problematic. Our findings suggest that even strong LLMs struggle with steerability, and existing alignment strategies may be insufficient. We open-source our steerability evaluation framework at https://github.com/MLD3/steerability.

Shengpu Tang, Jenna Wiens

Reinforcement learning (RL) can be used to learn treatment policies and aid decision making in healthcare. However, given the need for generalization over complex state/action spaces, the incorporation of function approximators (e.g., deep neural networks) requires model selection to reduce overfitting and improve policy performance at deployment. Yet a standard validation pipeline for model selection requires running a learned policy in the actual environment, which is often infeasible in a healthcare setting. In this work, we investigate a model selection pipeline for offline RL that relies on off-policy evaluation (OPE) as a proxy for validation performance. We present an in-depth analysis of popular OPE methods, highlighting the additional hyperparameters and computational requirements (fitting/inference of auxiliary models) when used to rank a set of candidate policies. We compare the utility of different OPE methods as part of the model selection pipeline in the context of learning to treat patients with sepsis. Among all the OPE methods we considered, fitted Q evaluation (FQE) consistently leads to the best validation ranking, but at a high computational cost. To balance this trade-off between accuracy of ranking and computational efficiency, we propose a simple two-stage approach to accelerate model selection by avoiding potentially unnecessary computation. Our work serves as a practical guide for offline RL model selection and can help RL practitioners select policies using real-world datasets. To facilitate reproducibility and future extensions, the code accompanying this paper is available online at https://github.com/MLD3/OfflineRL_ModelSelection.

Winston Chen, Michael W. Sjoding, Jenna Wiens

AI models are often evaluated based on their ability to predict the outcome of interest. However, in many AI for social impact applications, the presence of an intervention that affects the outcome can bias the evaluation. Randomized controlled trials (RCTs) randomly assign interventions, allowing data from the control group to be used for unbiased model evaluation. However, this approach is inefficient because it ignores data from the treatment group. Given the complexity and cost often associated with RCTs, making the most use of the data is essential. Thus, we investigate model evaluation strategies that leverage all data from an RCT. First, we theoretically quantify the estimation bias that arises from naïvely aggregating performance estimates from treatment and control groups and derive the condition under which this bias leads to incorrect model selection. Leveraging these theoretical insights, we propose nuisance parameter weighting (NPW), an unbiased model evaluation approach that reweights data from the treatment group to mimic the distributions of samples that would or would not experience the outcome under no intervention. Using synthetic and real-world datasets, we demonstrate that our proposed evaluation approach consistently yields better model selection than the standard approach, which ignores data from the treatment group, across various intervention effect and sample size settings. Our contribution represents a meaningful step towards more efficient model evaluation in real-world contexts.

Michael Ito, Jiong Zhu, Dexiong Chen et al.

In this work, we theoretically demonstrate that current graph positional encodings (PEs) are not beneficial and could potentially hurt performance in tasks involving heterophilous graphs, where nodes that are close tend to have different labels. This limitation is critical as many real-world networks exhibit heterophily, and even highly homophilous graphs can contain local regions of strong heterophily. To address this limitation, we propose Learnable Laplacian Positional Encodings (LLPE), a new PE that leverages the full spectrum of the graph Laplacian, enabling them to capture graph structure on both homophilous and heterophilous graphs. Theoretically, we prove LLPE's ability to approximate a general class of graph distances and demonstrate its generalization properties. Empirically, our evaluation on 12 benchmarks demonstrates that LLPE improves accuracy across a variety of GNNs, including graph transformers, by up to 35% and 14% on synthetic and real-world graphs, respectively. Going forward, our work represents a significant step towards developing PEs that effectively capture complex structures in heterophilous graphs.

Trenton Chang, Lindsay Warrenburg, Sae-Hwan Park et al.

In many settings, machine learning models may be used to inform decisions that impact individuals or entities who interact with the model. Such entities, or agents, may game model decisions by manipulating their inputs to the model to obtain better outcomes and maximize some utility. We consider a multi-agent setting where the goal is to identify the "worst offenders:" agents that are gaming most aggressively. However, identifying such agents is difficult without knowledge of their utility function. Thus, we introduce a framework in which each agent's tendency to game is parameterized via a scalar. We show that this gaming parameter is only partially identifiable. By recasting the problem as a causal effect estimation problem where different agents represent different "treatments," we prove that a ranking of all agents by their gaming parameters is identifiable. We present empirical results in a synthetic data study validating the usage of causal effect estimation for gaming detection and show in a case study of diagnosis coding behavior in the U.S. that our approach highlights features associated with gaming.

Michael Ito, Danai Koutra, Jenna Wiens

Homophily, as a measure, has been critical to increasing our understanding of graph neural networks (GNNs). However, to date this measure has only been analyzed in the context of static graphs. In our work, we explore homophily in dynamic settings. Focusing on graph convolutional networks (GCNs), we demonstrate theoretically that in dynamic settings, current GCN discriminative performance is characterized by the probability that a node's future label is the same as its neighbors' current labels. Based on this insight, we propose dynamic homophily, a new measure of homophily that applies in the dynamic setting. This new measure correlates with GNN discriminative performance and sheds light on how to potentially design more powerful GNNs for dynamic graphs. Leveraging a variety of dynamic node classification datasets, we demonstrate that popular GNNs are not robust to low dynamic homophily. Going forward, our work represents an important step towards understanding homophily and GNN performance in dynamic node classification.

Aishwarya Mandyam, Shengpu Tang, Jiayu Yao et al.

Off-policy evaluation (OPE) provides safety guarantees by estimating the performance of a policy before deployment. Recent work introduced IS+, an importance sampling (IS) estimator that uses expert-annotated counterfactual samples to improve behavior dataset coverage. However, IS estimators are known to have high variance; furthermore, the performance of IS+ deteriorates when annotations are imperfect. In this work, we propose a family of OPE estimators inspired by the doubly robust (DR) principle. A DR estimator combines IS with a reward model estimate, known as the direct method (DM), and offers favorable statistical guarantees. We propose three strategies for incorporating counterfactual annotations into a DR-inspired estimator and analyze their properties under various realistic settings. We prove that using imperfect annotations in the DM part of the estimator best leverages the annotations, as opposed to using them in the IS part. To support our theoretical findings, we evaluate the proposed estimators in three contextual bandit environments. Our empirical results show that when the reward model is misspecified and the annotations are imperfect, it is most beneficial to use the annotations only in the DM portion of a DR estimator. Based on these theoretical and empirical insights, we provide a practical guide for using counterfactual annotations in different realistic settings.

Markus Krimmel, Jenna Wiens, Karsten Borgwardt et al.

Graph generative models often face a critical trade-off between learning complex distributions and achieving fast generation speed. We introduce Autoregressive Noisy Filtration Modeling (ANFM), a novel approach that addresses both challenges. ANFM leverages filtration, a concept from topological data analysis, to transform graphs into short sequences of monotonically increasing subgraphs. This formulation extends the sequence families used in previous autoregressive models. To learn from these sequences, we propose a novel autoregressive graph mixer model. Our experiments suggest that exposure bias might represent a substantial hurdle in autoregressive graph generation and we introduce two mitigation strategies to address it: noise augmentation and a reinforcement learning approach. Incorporating these techniques leads to substantial performance gains, making ANFM competitive with state-of-the-art diffusion models across diverse synthetic and real-world datasets. Notably, ANFM produces remarkably short sequences, achieving a 100-fold speedup in generation time compared to diffusion models. This work marks a significant step toward high-throughput graph generation.

Michael Ito, Danai Koutra, Jenna Wiens

Random walk neural networks (RWNNs) have emerged as a promising approach for graph representation learning, leveraging recent advances in sequence models to process random walks. However, under realistic sampling constraints, RWNNs often fail to capture global structure even in small graphs due to incomplete node and edge coverage, limiting their expressivity. To address this, we propose \textit{random search neural networks} (RSNNs), which operate on random searches, each of which guarantees full node coverage. Theoretically, we demonstrate that in sparse graphs, only $O(\log |V|)$ searches are needed to achieve full edge coverage, substantially reducing sampling complexity compared to the $O(|V|)$ walks required by RWNNs (assuming walk lengths scale with graph size). Furthermore, when paired with universal sequence models, RSNNs are universal approximators. We lastly show RSNNs are probabilistically invariant to graph isomorphisms, ensuring their expectation is an isomorphism-invariant graph function. Empirically, RSNNs consistently outperform RWNNs on molecular and protein benchmarks, achieving comparable or superior performance with up to 16$\times$ fewer sampled sequences. Our work bridges theoretical and practical advances in random walk based approaches, offering an efficient and expressive framework for learning on sparse graphs.

Sarah Jabbour, David Fouhey, Nikola Banovic et al.

AI has the potential to augment human decision making. However, even high-performing models can produce inaccurate predictions when deployed. These inaccuracies, combined with automation bias, where humans overrely on AI predictions, can result in worse decisions. Selective prediction, in which potentially unreliable model predictions are hidden from users, has been proposed as a solution. This approach assumes that when AI abstains and informs the user so, humans make decisions as they would without AI involvement. To test this assumption, we study the effects of selective prediction on human decisions in a clinical context. We conducted a user study of 259 clinicians tasked with diagnosing and treating hospitalized patients. We compared their baseline performance without any AI involvement to their AI-assisted accuracy with and without selective prediction. Our findings indicate that selective prediction mitigates the negative effects of inaccurate AI in terms of decision accuracy. Compared to no AI assistance, clinician accuracy declined when shown inaccurate AI predictions (66% [95% CI: 56%-75%] vs. 56% [95% CI: 46%-66%]), but recovered under selective prediction (64% [95% CI: 54%-73%]). However, while selective prediction nearly maintains overall accuracy, our results suggest that it alters patterns of mistakes: when informed the AI abstains, clinicians underdiagnose (18% increase in missed diagnoses) and undertreat (35% increase in missed treatments) compared to no AI input at all. Our findings underscore the importance of empirically validating assumptions about how humans engage with AI within human-AI systems.

Dylan Zapzalka, Trenton Chang, Lindsay Warrenburg et al.

In settings where ML models are used to inform the allocation of resources, agents affected by the allocation decisions might have an incentive to strategically change their features to secure better outcomes. While prior work has studied strategic responses broadly, disentangling misreporting from genuine modification remains a fundamental challenge. In this paper, we propose a causally-motivated approach to identify and quantify how much an agent misreports on average by distinguishing deceptive changes in their features from genuine modification. Our key insight is that, unlike genuine modification, misreported features do not causally affect downstream variables (i.e., causal descendants). We exploit this asymmetry by comparing the causal effect of misreported features on their causal descendants as derived from manipulated datasets against those from unmanipulated datasets. We formally prove identifiability of the misreporting rate and characterize the variance of our estimator. We empirically validate our theoretical results using a semi-synthetic and real Medicare dataset with misreported data, demonstrating that our approach can be employed to identify misreporting in real-world scenarios.

Winston Chen, Trenton Chang, Jenna Wiens

Estimates of heterogeneous treatment assignment effects can inform treatment decisions. Under the presence of non-adherence (e.g., patients do not adhere to their assigned treatment), both the standard backdoor adjustment (SBD) and the conditional front-door adjustment (CFD) can recover unbiased estimates of the treatment assignment effects. However, the estimation variance of these approaches may vary widely across settings, which remains underexplored in the literature. In this work, we demonstrate theoretically and empirically that CFD yields lower-variance estimates than SBD when the true effect of treatment assignment is small (i.e., assigning an intervention leads to small changes in patients' future outcome). Additionally, since CFD requires estimating multiple nuisance parameters, we introduce LobsterNet, a multi-task neural network that implements CFD with joint modeling of the nuisance parameters. Empirically, LobsterNet reduces estimation error across several semi-synthetic and real-world datasets compared to baselines. Our findings suggest CFD with shared nuisance parameter modeling can improve treatment assignment effect estimation under non-adherence.

Trenton Chang, Jenna Wiens

Selective labels occur when label observations are subject to a decision-making process; e.g., diagnoses that depend on the administration of laboratory tests. We study a clinically-inspired selective label problem called disparate censorship, where labeling biases vary across subgroups and unlabeled individuals are imputed as "negative" (i.e., no diagnostic test = no illness). Machine learning models naively trained on such labels could amplify labeling bias. Inspired by causal models of selective labels, we propose Disparate Censorship Expectation-Maximization (DCEM), an algorithm for learning in the presence of disparate censorship. We theoretically analyze how DCEM mitigates the effects of disparate censorship on model performance. We validate DCEM on synthetic data, showing that it improves bias mitigation (area between ROC curves) without sacrificing discriminative performance (AUC) compared to baselines. We achieve similar results in a sepsis classification task using clinical data.

Shengpu Tang, Maggie Makar, Michael W. Sjoding et al.

Many reinforcement learning (RL) applications have combinatorial action spaces, where each action is a composition of sub-actions. A standard RL approach ignores this inherent factorization structure, resulting in a potential failure to make meaningful inferences about rarely observed sub-action combinations; this is particularly problematic for offline settings, where data may be limited. In this work, we propose a form of linear Q-function decomposition induced by factored action spaces. We study the theoretical properties of our approach, identifying scenarios where it is guaranteed to lead to zero bias when used to approximate the Q-function. Outside the regimes with theoretical guarantees, we show that our approach can still be useful because it leads to better sample efficiency without necessarily sacrificing policy optimality, allowing us to achieve a better bias-variance trade-off. Across several offline RL problems using simulators and real-world datasets motivated by healthcare, we demonstrate that incorporating factored action spaces into value-based RL can result in better-performing policies. Our approach can help an agent make more accurate inferences within underexplored regions of the state-action space when applying RL to observational datasets.

Sarah Jabbour, David Fouhey, Ella Kazerooni et al.

Objective: When patients develop acute respiratory failure, accurately identifying the underlying etiology is essential for determining the best treatment. However, differentiating between common medical diagnoses can be challenging in clinical practice. Machine learning models could improve medical diagnosis by aiding in the diagnostic evaluation of these patients. Materials and Methods: Machine learning models were trained to predict the common causes of acute respiratory failure (pneumonia, heart failure, and/or COPD). Models were trained using chest radiographs and clinical data from the electronic health record (EHR) and applied to an internal and external cohort. Results: The internal cohort of 1,618 patients included 508 (31%) with pneumonia, 363 (22%) with heart failure, and 137 (8%) with COPD based on physician chart review. A model combining chest radiographs and EHR data outperformed models based on each modality alone. Models had similar or better performance compared to a randomly selected physician reviewer. For pneumonia, the combined model area under the receiver operating characteristic curve (AUROC) was 0.79 (0.77-0.79), image model AUROC was 0.74 (0.72-0.75), and EHR model AUROC was 0.74 (0.70-0.76). For heart failure, combined: 0.83 (0.77-0.84), image: 0.80 (0.71-0.81), and EHR: 0.79 (0.75-0.82). For COPD, combined: AUROC = 0.88 (0.83-0.91), image: 0.83 (0.77-0.89), and EHR: 0.80 (0.76-0.84). In the external cohort, performance was consistent for heart failure and increased for COPD, but declined slightly for pneumonia. Conclusions: Machine learning models combining chest radiographs and EHR data can accurately differentiate between common causes of acute respiratory failure. Further work is needed to determine how these models could act as a diagnostic aid to clinicians in clinical settings.

Erkin Ötleş, Jeeheh Oh, Benjamin Li et al.

Once integrated into clinical care, patient risk stratification models may perform worse compared to their retrospective performance. To date, it is widely accepted that performance will degrade over time due to changes in care processes and patient populations. However, the extent to which this occurs is poorly understood, in part because few researchers report prospective validation performance. In this study, we compare the 2020-2021 ('20-'21) prospective performance of a patient risk stratification model for predicting healthcare-associated infections to a 2019-2020 ('19-'20) retrospective validation of the same model. We define the difference in retrospective and prospective performance as the performance gap. We estimate how i) "temporal shift", i.e., changes in clinical workflows and patient populations, and ii) "infrastructure shift", i.e., changes in access, extraction and transformation of data, both contribute to the performance gap. Applied prospectively to 26,864 hospital encounters during a twelve-month period from July 2020 to June 2021, the model achieved an area under the receiver operating characteristic curve (AUROC) of 0.767 (95% confidence interval (CI): 0.737, 0.801) and a Brier score of 0.189 (95% CI: 0.186, 0.191). Prospective performance decreased slightly compared to '19-'20 retrospective performance, in which the model achieved an AUROC of 0.778 (95% CI: 0.744, 0.815) and a Brier score of 0.163 (95% CI: 0.161, 0.165). The resulting performance gap was primarily due to infrastructure shift and not temporal shift. So long as we continue to develop and validate models using data stored in large research data warehouses, we must consider differences in how and when data are accessed, measure how these differences may affect prospective performance, and work to mitigate those differences.

Jiaxuan Wang, Jenna Wiens, Scott Lundberg

Many existing approaches for estimating feature importance are problematic because they ignore or hide dependencies among features. A causal graph, which encodes the relationships among input variables, can aid in assigning feature importance. However, current approaches that assign credit to nodes in the causal graph fail to explain the entire graph. In light of these limitations, we propose Shapley Flow, a novel approach to interpreting machine learning models. It considers the entire causal graph, and assigns credit to \textit{edges} instead of treating nodes as the fundamental unit of credit assignment. Shapley Flow is the unique solution to a generalization of the Shapley value axioms to directed acyclic graphs. We demonstrate the benefit of using Shapley Flow to reason about the impact of a model's input on its output. In addition to maintaining insights from existing approaches, Shapley Flow extends the flat, set-based, view prevalent in game theory based explanation methods to a deeper, \textit{graph-based}, view. This graph-based view enables users to understand the flow of importance through a system, and reason about potential interventions.

Sarah Jabbour, David Fouhey, Ella Kazerooni et al.

While deep learning has shown promise in improving the automated diagnosis of disease based on chest X-rays, deep networks may exhibit undesirable behavior related to shortcuts. This paper studies the case of spurious class skew in which patients with a particular attribute are spuriously more likely to have the outcome of interest. For instance, clinical protocols might lead to a dataset in which patients with pacemakers are disproportionately likely to have congestive heart failure. This skew can lead to models that take shortcuts by heavily relying on the biased attribute. We explore this problem across a number of attributes in the context of diagnosing the cause of acute hypoxemic respiratory failure. Applied to chest X-rays, we show that i) deep nets can accurately identify many patient attributes including sex (AUROC = 0.96) and age (AUROC >= 0.90), ii) they tend to exploit correlations between such attributes and the outcome label when learning to predict a diagnosis, leading to poor performance when such correlations do not hold in the test population (e.g., everyone in the test set is male), and iii) a simple transfer learning approach is surprisingly effective at preventing the shortcut and promoting good generalization performance. On the task of diagnosing congestive heart failure based on a set of chest X-rays skewed towards older patients (age >= 63), the proposed approach improves generalization over standard training from 0.66 (95% CI: 0.54-0.77) to 0.84 (95% CI: 0.73-0.92) AUROC. While simple, the proposed approach has the potential to improve the performance of models across populations by encouraging reliance on clinically relevant manifestations of disease, i.e., those that a clinician would use to make a diagnosis.

Ian Fox, Joyce Lee, Rodica Pop-Busui et al.

People with type 1 diabetes (T1D) lack the ability to produce the insulin their bodies need. As a result, they must continually make decisions about how much insulin to self-administer to adequately control their blood glucose levels. Longitudinal data streams captured from wearables, like continuous glucose monitors, can help these individuals manage their health, but currently the majority of the decision burden remains on the user. To relieve this burden, researchers are working on closed-loop solutions that combine a continuous glucose monitor and an insulin pump with a control algorithm in an `artificial pancreas.' Such systems aim to estimate and deliver the appropriate amount of insulin. Here, we develop reinforcement learning (RL) techniques for automated blood glucose control. Through a series of experiments, we compare the performance of different deep RL approaches to non-RL approaches. We highlight the flexibility of RL approaches, demonstrating how they can adapt to new individuals with little additional data. On over 2.1 million hours of data from 30 simulated patients, our RL approach outperforms baseline control algorithms: leading to a decrease in median glycemic risk of nearly 50% from 8.34 to 4.24 and a decrease in total time hypoglycemic of 99.8%, from 4,610 days to 6. Moreover, these approaches are able to adapt to predictable meal times (decreasing average risk by an additional 24% as meals increase in predictability). This work demonstrates the potential of deep RL to help people with T1D manage their blood glucose levels without requiring expert knowledge. All of our code is publicly available, allowing for replication and extension.

Shengpu Tang, Aditya Modi, Michael W. Sjoding et al.

Standard reinforcement learning (RL) aims to find an optimal policy that identifies the best action for each state. However, in healthcare settings, many actions may be near-equivalent with respect to the reward (e.g., survival). We consider an alternative objective -- learning set-valued policies to capture near-equivalent actions that lead to similar cumulative rewards. We propose a model-free algorithm based on temporal difference learning and a near-greedy heuristic for action selection. We analyze the theoretical properties of the proposed algorithm, providing optimality guarantees and demonstrate our approach on simulated environments and a real clinical task. Empirically, the proposed algorithm exhibits good convergence properties and discovers meaningful near-equivalent actions. Our work provides theoretical, as well as practical, foundations for clinician/human-in-the-loop decision making, in which humans (e.g., clinicians, patients) can incorporate additional knowledge (e.g., side effects, patient preference) when selecting among near-equivalent actions.

Jiaxuan Wang, Jenna Wiens

In the context of stochastic gradient descent(SGD) and adaptive moment estimation (Adam),researchers have recently proposed optimization techniques that transition from Adam to SGD with the goal of improving both convergence and generalization performance. However, precisely how each approach trades off early progress and generalization is not well understood; thus, it is unclear when or even if, one should transition from one approach to the other. In this work, by first studying the convex setting, we identify potential contributors to observed differences in performance between SGD and Adam. In particular,we provide theoretical insights for when and why Adam outperforms SGD and vice versa. We ad-dress the performance gap by adapting a single global learning rate for SGD, which we refer to as AdaSGD. We justify this proposed approach with empirical analyses in non-convex settings. On several datasets that span three different domains,we demonstrate how AdaSGD combines the benefits of both SGD and Adam, eliminating the need for approaches that transition from Adam to SGD.

Ian Fox, Jenna Wiens

We introduce advocacy learning, a novel supervised training scheme for attention-based classification problems. Advocacy learning relies on a framework consisting of two connected networks: 1) $N$ Advocates (one for each class), each of which outputs an argument in the form of an attention map over the input, and 2) a Judge, which predicts the class label based on these arguments. Each Advocate produces a class-conditional representation with the goal of convincing the Judge that the input example belongs to their class, even when the input belongs to a different class. Applied to several different classification tasks, we show that advocacy learning can lead to small improvements in classification accuracy over an identical supervised baseline. Though a series of follow-up experiments, we analyze when and how such class-conditional representations improve discriminative performance. Though somewhat counter-intuitive, a framework in which subnetworks are trained to competitively provide evidence in support of their class shows promise, in many cases performing on par with standard learning approaches. This provides a foundation for further exploration into competition and class-conditional representations in supervised learning.

Tian Bao, Brooke N. Klatt, Susan L. Whitney et al.

Compared to in-clinic balance training, in-home training is not as effective. This is, in part, due to the lack of feedback from physical therapists (PTs). Here, we analyze the feasibility of using trunk sway data and machine learning (ML) techniques to automatically evaluate balance, providing accurate assessments outside of the clinic. We recruited sixteen participants to perform standing balance exercises. For each exercise, we recorded trunk sway data and had a PT rate balance performance on a scale of 1 to 5. The rating scale was adapted from the Functional Independence Measure. From the trunk sway data, we extracted a 61-dimensional feature vector representing performance of each exercise. Given these labeled data, we trained a multi-class support vector machine (SVM) to map trunk sway features to PT ratings. Evaluated in a leave-one-participant-out scheme, the model achieved a classification accuracy of 82%. Compared to participant self-assessment ratings, the SVM outputs were significantly closer to PT ratings. The results of this pilot study suggest that in the absence of PTs, ML techniques can provide accurate assessments during standing balance exercises. Such automated assessments could reduce PT consultation time and increase user compliance outside of the clinic.

Jeeheh Oh, Jiaxuan Wang, Shengpu Tang et al.

Recurrent neural networks (RNNs) are commonly applied to clinical time-series data with the goal of learning patient risk stratification models. Their effectiveness is due, in part, to their use of parameter sharing over time (i.e., cells are repeated hence the name recurrent). We hypothesize, however, that this trait also contributes to the increased difficulty such models have with learning relationships that change over time. Conditional shift, i.e., changes in the relationship between the input X and the output y, arises when risk factors associated with the event of interest change over the course of a patient admission. While in theory, RNNs and gated RNNs (e.g., LSTMs) in particular should be capable of learning time-varying relationships, when training data are limited, such models often fail to accurately capture these dynamics. We illustrate the advantages and disadvantages of complete parameter sharing (RNNs) by comparing an LSTM with shared parameters to a sequential architecture with time-varying parameters on prediction tasks involving three clinically-relevant outcomes: acute respiratory failure (ARF), shock, and in-hospital mortality. In experiments using synthetic data, we demonstrate how parameter sharing in LSTMs leads to worse performance in the presence of conditional shift. To improve upon the dichotomy between complete parameter sharing and no parameter sharing, we propose a novel RNN formulation based on a mixture model in which we relax parameter sharing over time. The proposed method outperforms standard LSTMs and other state-of-the-art baselines across all tasks. In settings with limited data, relaxed parameter sharing can lead to improved patient risk stratification performance.

Eli Sherman, Hitinder Gurm, Ulysses Balis et al.

In healthcare, patient risk stratification models are often learned using time-series data extracted from electronic health records. When extracting data for a clinical prediction task, several formulations exist, depending on how one chooses the time of prediction and the prediction horizon. In this paper, we show how the formulation can greatly impact both model performance and clinical utility. Leveraging a publicly available ICU dataset, we consider two clinical prediction tasks: in-hospital mortality, and hypokalemia. Through these case studies, we demonstrate the necessity of evaluating models using an outcome-independent reference point, since choosing the time of prediction relative to the event can result in unrealistic performance. Further, an outcome-independent scheme outperforms an outcome-dependent scheme on both tasks (In-Hospital Mortality AUROC .882 vs. .831; Serum Potassium: AUROC .829 vs. .740) when evaluated on test sets that mimic real-world use.

Jeeheh Oh, Jiaxuan Wang, Jenna Wiens

Recently, researchers have started applying convolutional neural networks (CNNs) with one-dimensional convolutions to clinical tasks involving time-series data. This is due, in part, to their computational efficiency, relative to recurrent neural networks and their ability to efficiently exploit certain temporal invariances, (e.g., phase invariance). However, it is well-established that clinical data may exhibit many other types of invariances (e.g., scaling). While preprocessing techniques, (e.g., dynamic time warping) may successfully transform and align inputs, their use often requires one to identify the types of invariances in advance. In contrast, we propose the use of Sequence Transformer Networks, an end-to-end trainable architecture that learns to identify and account for invariances in clinical time-series data. Applied to the task of predicting in-hospital mortality, our proposed approach achieves an improvement in the area under the receiver operating characteristic curve (AUROC) relative to a baseline CNN (AUROC=0.851 vs. AUROC=0.838). Our results suggest that a variety of valuable invariances can be learned directly from the data.

Pascal Sturmfels, Saige Rutherford, Mike Angstadt et al.

Given the wide success of convolutional neural networks (CNNs) applied to natural images, researchers have begun to apply them to neuroimaging data. To date, however, exploration of novel CNN architectures tailored to neuroimaging data has been limited. Several recent works fail to leverage the 3D structure of the brain, instead treating the brain as a set of independent 2D slices. Approaches that do utilize 3D convolutions rely on architectures developed for object recognition tasks in natural 2D images. Such architectures make assumptions about the input that may not hold for neuroimaging. For example, existing architectures assume that patterns in the brain exhibit translation invariance. However, a pattern in the brain may have different meaning depending on where in the brain it is located. There is a need to explore novel architectures that are tailored to brain images. We present two simple modifications to existing CNN architectures based on brain image structure. Applied to the task of brain age prediction, our network achieves a mean absolute error (MAE) of 1.4 years and trains 30% faster than a CNN baseline that achieves a MAE of 1.6 years. Our results suggest that lessons learned from developing models on natural images may not directly transfer to neuroimaging tasks. Instead, there remains a large space of unexplored questions regarding model development in this area, whose answers may differ from conventional wisdom.

Ian Fox, Lynn Ang, Mamta Jaiswal et al.

In many forecasting applications, it is valuable to predict not only the value of a signal at a certain time point in the future, but also the values leading up to that point. This is especially true in clinical applications, where the future state of the patient can be less important than the patient's overall trajectory. This requires multi-step forecasting, a forecasting variant where one aims to predict multiple values in the future simultaneously. Standard methods to accomplish this can propagate error from prediction to prediction, reducing quality over the long term. In light of these challenges, we propose multi-output deep architectures for multi-step forecasting in which we explicitly model the distribution of future values of the signal over a prediction horizon. We apply these techniques to the challenging and clinically relevant task of blood glucose forecasting. Through a series of experiments on a real-world dataset consisting of 550K blood glucose measurements, we demonstrate the effectiveness of our proposed approaches in capturing the underlying signal dynamics. Compared to existing shallow and deep methods, we find that our proposed approaches improve performance individually and capture complementary information, leading to a large improvement over the baseline when combined (4.87 vs. 5.31 absolute percentage error (APE)). Overall, the results suggest the efficacy of our proposed approach in predicting blood glucose level and multi-step forecasting more generally.

Jiaxuan Wang, Ian Fox, Jonathan Skaza et al.

During the 2017 NBA playoffs, Celtics coach Brad Stevens was faced with a difficult decision when defending against the Cavaliers: "Do you double and risk giving up easy shots, or stay at home and do the best you can?" It's a tough call, but finding a good defensive strategy that effectively incorporates doubling can make all the difference in the NBA. In this paper, we analyze double teaming in the NBA, quantifying the trade-off between risk and reward. Using player trajectory data pertaining to over 643,000 possessions, we identified when the ball handler was double teamed. Given these data and the corresponding outcome (i.e., was the defense successful), we used deep reinforcement learning to estimate the quality of the defensive actions. We present qualitative and quantitative results summarizing our learned defensive strategy for defending. We show that our policy value estimates are predictive of points per possession and win percentage. Overall, the proposed framework represents a step toward a more comprehensive understanding of defensive strategies in the NBA.

Dev Goyal, Zeeshan Syed, Jenna Wiens

Longitudinal patient data has the potential to improve clinical risk stratification models for disease. However, chronic diseases that progress slowly over time are often heterogeneous in their clinical presentation. Patients may progress through disease stages at varying rates. This leads to pathophysiological misalignment over time, making it difficult to consistently compare patients in a clinically meaningful way. Furthermore, patients present clinically for the first time at different stages of disease. This eliminates the possibility of simply aligning patients based on their initial presentation. Finally, patient data may be sampled at different rates due to differences in schedules or missed visits. To address these challenges, we propose a robust measure of patient similarity based on subsequence alignment. Compared to global alignment techniques that do not account for pathophysiological misalignment, focusing on the most relevant subsequences allows for an accurate measure of similarity between patients. We demonstrate the utility of our approach in settings where longitudinal data, while useful, are limited and lack a clear temporal alignment for comparison. Applied to the task of stratifying patients for risk of progression to probable Alzheimer's Disease, our approach outperforms models that use only snapshot data (AUROC of 0.839 vs. 0.812) and models that use global alignment techniques (AUROC of 0.822). Our results support the hypothesis that patients' trajectories are useful for quantifying inter-patient similarities and that using subsequence matching and can help account for heterogeneity and misalignment in longitudinal data.

Jiaxuan Wang, Jeeheh Oh, Haozhu Wang et al.

In many settings, it is important that a model be capable of providing reasons for its predictions (i.e., the model must be interpretable). However, the model's reasoning may not conform with well-established knowledge. In such cases, while interpretable, the model lacks \textit{credibility}. In this work, we formally define credibility in the linear setting and focus on techniques for learning models that are both accurate and credible. In particular, we propose a regularization penalty, expert yielded estimates (EYE), that incorporates expert knowledge about well-known relationships among covariates and the outcome of interest. We give both theoretical and empirical results comparing our proposed method to several other regularization techniques. Across a range of settings, experiments on both synthetic and real data show that models learned using the EYE penalty are significantly more credible than those learned using other penalties. Applied to a large-scale patient risk stratification task, our proposed technique results in a model whose top features overlap significantly with known clinical risk factors, while still achieving good predictive performance.

Ian Fox, Lynn Ang, Mamta Jaiswal et al.

Motifs are a powerful tool for analyzing physiological waveform data. Standard motif methods, however, ignore important contextual information (e.g., what the patient was doing at the time the data were collected). We hypothesize that these additional contextual data could increase the utility of motifs. Thus, we propose an extension to motifs, contextual motifs, that incorporates context. Recognizing that, oftentimes, context may be unobserved or unavailable, we focus on methods to jointly infer motifs and context. Applied to both simulated and real physiological data, our proposed approach improves upon existing motif methods in terms of the discriminative utility of the discovered motifs. In particular, we discovered contextual motifs in continuous glucose monitor (CGM) data collected from patients with type 1 diabetes. Compared to their contextless counterparts, these contextual motifs led to better predictions of hypo- and hyperglycemic events. Our results suggest that even when inferred, context is useful in both a long- and short-term prediction horizon when processing and interpreting physiological waveform data.