A structural characterization of shortcut features for prediction.

With the rising use of machine learning for healthcare applications, practitioners are increasingly confronted with the limitations of prediction models that are trained in one setting but meant to be deployed in several others. One recently identified limitation is so-called shortcut learning, whereby a model learns to associate features with the prediction target that do not maintain their relationship across settings. Famously, the watermark on chest x-rays has been demonstrated to be an instance of a shortcut feature. In this viewpoint, we attempt to give a structural characterization of shortcut features in terms of causal DAGs. This is the first attempt at defining shortcut features in terms of their causal relationship with a model's prediction target.

Introduction

Data analyses can be carried out for 3 tasks: description, prediction, and counterfactual prediction; the latter is used for causal inference [1]. Each of these tasks requires the combination of data with different types of external or expert knowledge. For example, when performing a causal inference task, a core piece of expert knowledge is the specification of the set of variables to adjust for confounding [2]. Domain experts may select those adjustment variables after representing their knowledge about the underlying causal structure using causal directed acyclic graphs (DAGs) [3].

In contrast, the role of expert knowledge about the causal structure is less clear for prediction tasks. The goal of prediction is to learn a mapping from a set of input features to the prediction target. For example, the pixels from a chest X-ray (the features) may be used to predict the presence of a disease (the target). Recently, prediction models based on machine learning have demonstrated impressive results on a broad set of problems that have historically been challenging prediction tasks [4-10]. The success in machine learning has largely been driven by deep learning [11] techniques, which are extremely flexible models that impose very few assumptions and limited amounts of explicit expert knowledge [8]. The canonical deep learning approach is to include as many features as possible and to learn, solely from data, which combinations and transformations of these variables result in the most accurate prediction model [12]. Thus, it seems that the role of expert knowledge has an increasingly small role to play in prediction modeling. However, recent work has brought into question the wisdom of the purely data-driven paradigm in prediction modeling. In this viewpoint, we describe the role of expert knowledge in prediction using causal DAGs.

An example of a shortcut feature

Suppose that we train a deep learning model to predict whether a patient has COVID-19 (the prediction target) using chest X-rays (the input features) from a healthcare system of 4 hospitals. The goal is that the model uses pathophysiological markers such as lung opacity, like a human physician, to create a clinical decision support tool that can be deployed in the 4 hospitals as well as in additional hospitals that did not provide training data.

However, the model might find a shortcut to predict disease using the training data but without using pathophysiological markers. Because some hospitals are more likely to treat patients with COVID-19, the X-rays from those hospitals are more likely to correspond to individuals with COVID-19 than the X-rays from other hospitals. Therefore, like in a recent study, a deep learning model may associate the presence of a hospital-specific watermark in the top-right corner of each X-ray with a greater probability of COVID-19 diagnosis [13]. The watermark in the chest X-ray is an example of a predictor that has been referred to as a shortcut feature [13, 14] in the machine learning literature. While the presence of a watermark for a specific hospital does shift the relative probability of COVID-19, the predictive power of the watermark is of limited use when the deep learning model is applied to X-rays from new hospitals with different watermarks or no watermark at all.

The characterization and avoidance of shortcut features is an active area of research [15-28]. Here we argue that expert knowledge is required to define what a shortcut feature is and to distinguish between shortcut features and other features.

Three types of input features in prediction models

The causal DAG in Fig. 1A depicts the prediction target (COVID-19 diagnosis) and the input features (lung opacity and watermark from the X rays). The arrow from COVID-19 to lung opacity represents the causal effect of the infection that is mediated by changes in the lung, while the arrow from COVID-19 to watermark represents the causal effect of infection that is mediated by admitting hospital, and the absence of an arrow from COVID-19 to number of ribs represents the lack of effect of infection on the bones. We say that a path in a causal DAG is open when it only includes two types of nodes: colliders (or their descendants) that are conditioned on, or non-colliders that are not [29, 30]. Two variables connected by an open path are expected to be statistically associated.

Fig. 1

Causal DAG underlying the prediction of COVID-19 status from chest X-rays across two different contexts. The training context (left) includes 4 hospitals with different COVID-19 prevalence and watermarking patterns. The deployment context (right) includes a fifth hospital with no watermark on the X-ray

Using causal DAGs, we can consider 3 types of input features based on the paths that link them to the prediction target:

Non-shortcut features are connected with the prediction target exclusively through open paths that are conserved across the training and deployment settings, e.g., lung opacity.

Shortcut features are connected with the prediction target through one or more open paths that are not conserved across the training and deployment settings, e.g., watermarks in the X-rays from the original hospitals (which do not provide information on COVID-19 diagnosis in other hospitals).

Irrelevant features are not connected with the prediction target through any open paths in the training data, e.g., number of ribs.

Our classification is qualitative (because we do not provide a quantitative definition of “conserved path”) and informal, but it is sufficient to extract some conclusions about the role of expert knowledge for prediction tasks and a basis for the development of more formal approaches.

The role of expert knowledge

Expert knowledge is not required to conclude that irrelevant features (c) are not helpful to improve the prediction model. Though they may be statistically associated with the prediction target by chance in finite samples, irrelevant features (c) will be automatically discarded by the deep learning model given enough training data. In contrast, both non-shortcut (a) and shortcut (b) features will be incorporated into the predictive model, but the only way to distinguish between non-shortcut and shortcut features is the use of causal knowledge.

The classification of a feature as a shortcut feature depends on the set of deployment settings under consideration. Therefore, building a reliable prediction model requires that we restrict the model’s access to shortcut features relative to the expected deployment contexts. In the chest X-ray example, lung opacity is a non-shortcut feature (a) because we expect it to be affected by COVID-19, regardless of which hospital performed the X-ray. In contrast, the watermark is a shortcut feature (b) because it lies on an open path to the prediction target in the hospitals used to train the model, but this path no longer exists in other hospitals in which the model is deployed, since the deployment hospital uses a different or no watermark.

This failure of predictive models to distinguish features (a) and (b) reflects the model’s ignorance about the underlying causal structures in the training and deployment contexts. Therefore, model developers need to use expert knowledge to characterize the relevant set of deployment contexts, identify the shortcut features, and exclude the shortcut features from the data before the model is trained. In our example, model developers could accomplish this by cropping out all watermarks from the X-rays. Some shortcut features may be less obvious to experts or may be difficult to define concisely. For example, it has been shown that scanners from different manufacturers encode information in chest X-rays in slightly different ways [31]. This results in a nebulous scanner “signature”, invisible to the human eye, that can result in shortcut learning if different patient populations are scanned using machines from different companies [31].

Causal structures for shortcut features

In general, features (a) and (b) do not have to be causally affected by the prediction target. Instead, the open path between a shortcut feature and the outcome may have a variety of causal structures (see Fig. 2), where a subset of the open path differs between the training and deployment contexts. In the most extreme case, a subset of the arrows and nodes are present in one setting, but absent in the other. Figure 2 illustrates some possible causal structures for shortcut features.

Fig. 2

Causal DAGs in the training (top row) and deployment (bottom row) context. Panel (A) depicts the same causal structure as Fig. 1, panels (B) and (C) depict alternative causal structures for shortcut features

The left column shows the same causal structure that we examined in Fig. 1. The prediction target precedes the features temporally thus the direction of causation and prediction are opposite, also known as anti-causal prediction [32]. The central column shows a shortcut feature that shares a common cause with the prediction target in the training data but not in the deployment setting. The right column shows a shortcut feature that affects the prediction target in the training setting but not in the deployment setting. Figure 2 is not meant to be an exhaustive characterization of causal structures that permit shortcut learning, as we can imagine other structures with paths containing colliders between the shortcut feature and the prediction target that change between training and deployment.

Other views on shortcut features

Shortcut features have attracted a great deal of attention from research groups in the fields of statistics, machine learning, and causal inference. We offer a brief overview of related efforts and provide a contrast with the characterization we provide in this work.

Arjovsky et al. [15] introduced invariant risk minimization (IRM) for data collected across a set of pre-specified settings: the model is trained to minimize the maximum error across settings (i.e., a minimax objective) and, by leveraging the pre-specification of deployment settings, the IRM framework is expected to discard shortcut features even if the causal structure of shortcut features is not explicitly characterized.

Other work has defined shortcuts in the representations learned by deep learning models rather than in the features themselves. For example, Wang and Jordan [25] consider a setting in which the input features cause the prediction target Y and there is an unobserved common cause C of all features (Fig. 3). Under the representation learning framework, models attempt to learn a mapping from the input feature space to a lower-dimensional representation of each sample. In general, we would like this representation to be highly predictive of the outcome and free of shortcuts, which the authors operationalize by defining a measure of a representation’s sufficiency for the outcome. However, their notion of sufficiency does not apply to individual features but instead applies to the black box representation of the inputs. While this framework may enable the development of methods that optimize for representations with high sufficiency, it does not advance our reasoning about individual features, which is often the most useful level for reasoning about shortcuts in epidemiology.

Fig. 3

Causal DAG from Wang et al. [25]. The features X1, …, Xm cause the prediction target Y and share a common cause C

In the field of statistical transportability, previous work has been concerned with estimating the conditional distribution of Y given X in a target domain using the distribution learned in the training domain and limited data from the target domain [33-39]. Correa and Bareinboim [26] developed an approach for solving this task that relies on factorizing the target conditional distribution according to its underlying causal DAG using a novel formalism known as a c*-factors. It has yet to be shown if the conditions required by this approach from this work can be stated in terms of graphical conditions that can be easily read off a DAG without requiring c*-factors. However, shortcut features per our definition would violate these transportability conditions. We believe that a direct structural characterization of shortcuts at the level of individual features (as opposed to at the level of c*-factors) is the most useful for practical discussion and enables intuitive reasoning about the surrounding issues.

Conclusion

Shortcut features pose a significant challenge to the safe and reliable deployment of prediction models. Models that learn shortcut features are unreliable and have the potential to cause catastrophic errors if used in clinical decision-making. An explicitly causal characterization of shortcut features, as we proposed here, facilitates the incorporation of expert knowledge into prediction models and may guide future work on remedies to the problem.