Christopher Lee-Messer

Ali Mirzazadeh, Florian Dubost, Maxwell Pike et al. · stanford

Attention--or attribution--maps methods are methods designed to highlight regions of the model's input that were discriminative for its predictions. However, different attention maps methods can highlight different regions of the input, with sometimes contradictory explanations for a prediction. This effect is exacerbated when the training set is small. This indicates that either the model learned incorrect representations or that the attention maps methods did not accurately estimate the model's representations. We propose an unsupervised fine-tuning method that optimizes the consistency of attention maps and show that it improves both classification performance and the quality of attention maps. We propose an implementation for two state-of-the-art attention computation methods, Grad-CAM and Guided Backpropagation, which relies on an input masking technique. We also show results on Grad-CAM and Integrated Gradients in an ablation study. We evaluate this method on our own dataset of event detection in continuous video recordings of hospital patients aggregated and curated for this work. As a sanity check, we also evaluate the proposed method on PASCAL VOC and SVHN. With the proposed method, with small training sets, we achieve a 6.6 points lift of F1 score over the baselines on our video dataset, a 2.9 point lift of F1 score on PASCAL, and a 1.8 points lift of mean Intersection over Union over Grad-CAM for weakly supervised detection on PASCAL. Those improved attention maps may help clinicians better understand vision model predictions and ease the deployment of machine learning systems into clinical care. We share part of the code for this article at the following repository: https://github.com/alimirzazadeh/SemisupervisedAttention.

Sabri Eyuboglu, Maya Varma, Khaled Saab et al. · stanford

Machine learning models that achieve high overall accuracy often make systematic errors on important subsets (or slices) of data. Identifying underperforming slices is particularly challenging when working with high-dimensional inputs (e.g. images, audio), where important slices are often unlabeled. In order to address this issue, recent studies have proposed automated slice discovery methods (SDMs), which leverage learned model representations to mine input data for slices on which a model performs poorly. To be useful to a practitioner, these methods must identify slices that are both underperforming and coherent (i.e. united by a human-understandable concept). However, no quantitative evaluation framework currently exists for rigorously assessing SDMs with respect to these criteria. Additionally, prior qualitative evaluations have shown that SDMs often identify slices that are incoherent. In this work, we address these challenges by first designing a principled evaluation framework that enables a quantitative comparison of SDMs across 1,235 slice discovery settings in three input domains (natural images, medical images, and time-series data). Then, motivated by the recent development of powerful cross-modal representation learning approaches, we present Domino, an SDM that leverages cross-modal embeddings and a novel error-aware mixture model to discover and describe coherent slices. We find that Domino accurately identifies 36% of the 1,235 slices in our framework - a 12 percentage point improvement over prior methods. Further, Domino is the first SDM that can provide natural language descriptions of identified slices, correctly generating the exact name of the slice in 35% of settings.

Siyi Tang, Jared A. Dunnmon, Liangqiong Qu et al. · stanford

Multivariate biosignals are prevalent in many medical domains, such as electroencephalography, polysomnography, and electrocardiography. Modeling spatiotemporal dependencies in multivariate biosignals is challenging due to (1) long-range temporal dependencies and (2) complex spatial correlations between the electrodes. To address these challenges, we propose representing multivariate biosignals as time-dependent graphs and introduce GraphS4mer, a general graph neural network (GNN) architecture that improves performance on biosignal classification tasks by modeling spatiotemporal dependencies in biosignals. Specifically, (1) we leverage the Structured State Space architecture, a state-of-the-art deep sequence model, to capture long-range temporal dependencies in biosignals and (2) we propose a graph structure learning layer in GraphS4mer to learn dynamically evolving graph structures in the data. We evaluate our proposed model on three distinct biosignal classification tasks and show that GraphS4mer consistently improves over existing models, including (1) seizure detection from electroencephalographic signals, outperforming a previous GNN with self-supervised pre-training by 3.1 points in AUROC; (2) sleep staging from polysomnographic signals, a 4.1 points improvement in macro-F1 score compared to existing sleep staging models; and (3) 12-lead electrocardiogram classification, outperforming previous state-of-the-art models by 2.7 points in macro-F1 score.

Nandita Bhaskhar, Daniel L. Rubin, Christopher Lee-Messer · stanford

Continuous monitoring of trained ML models to determine when their predictions should and should not be trusted is essential for their safe deployment. Such a framework ought to be high-performing, explainable, post-hoc and actionable. We propose TRUST-LAPSE, a "mistrust" scoring framework for continuous model monitoring. We assess the trustworthiness of each input sample's model prediction using a sequence of latent-space embeddings. Specifically, (a) our latent-space mistrust score estimates mistrust using distance metrics (Mahalanobis distance) and similarity metrics (cosine similarity) in the latent-space and (b) our sequential mistrust score determines deviations in correlations over the sequence of past input representations in a non-parametric, sliding-window based algorithm for actionable continuous monitoring. We evaluate TRUST-LAPSE via two downstream tasks: (1) distributionally shifted input detection, and (2) data drift detection. We evaluate across diverse domains - audio and vision using public datasets and further benchmark our approach on challenging, real-world electroencephalograms (EEG) datasets for seizure detection. Our latent-space mistrust scores achieve state-of-the-art results with AUROCs of 84.1 (vision), 73.9 (audio), and 77.1 (clinical EEGs), outperforming baselines by over 10 points. We expose critical failures in popular baselines that remain insensitive to input semantic content, rendering them unfit for real-world model monitoring. We show that our sequential mistrust scores achieve high drift detection rates; over 90% of the streams show < 20% error for all domains. Through extensive qualitative and quantitative evaluations, we show that our mistrust scores are more robust and provide explainability for easy adoption into practice.

Khaled Saab, Siyi Tang, Mohamed Taha et al. · stanford

A major barrier to deploying healthcare AI models is their trustworthiness. One form of trustworthiness is a model's robustness across different subgroups: while existing models may exhibit expert-level performance on aggregate metrics, they often rely on non-causal features, leading to errors in hidden subgroups. To take a step closer towards trustworthy seizure onset detection from EEG, we propose to leverage annotations that are produced by healthcare personnel in routine clinical workflows -- which we refer to as workflow notes -- that include multiple event descriptions beyond seizures. Using workflow notes, we first show that by scaling training data to an unprecedented level of 68,920 EEG hours, seizure onset detection performance significantly improves (+12.3 AUROC points) compared to relying on smaller training sets with expensive manual gold-standard labels. Second, we reveal that our binary seizure onset detection model underperforms on clinically relevant subgroups (e.g., up to a margin of 6.5 AUROC points between pediatrics and adults), while having significantly higher false positives on EEG clips showing non-epileptiform abnormalities compared to any EEG clip (+19 FPR points). To improve model robustness to hidden subgroups, we train a multilabel model that classifies 26 attributes other than seizures, such as spikes, slowing, and movement artifacts. We find that our multilabel model significantly improves overall seizure onset detection performance (+5.9 AUROC points) while greatly improving performance among subgroups (up to +8.3 AUROC points), and decreases false positives on non-epileptiform abnormalities by 8 FPR points. Finally, we propose a clinical utility metric based on false positives per 24 EEG hours and find that our multilabel model improves this clinical utility metric by a factor of 2x across different clinical settings.

Siddharth Sharma, Florian Dubost, Christopher Lee-Messer et al.

In a clinical setting, epilepsy patients are monitored via video electroencephalogram (EEG) tests. A video EEG records what the patient experiences on videotape while an EEG device records their brainwaves. Currently, there are no existing automated methods for tracking the patient's location during a seizure, and video recordings of hospital patients are substantially different from publicly available video benchmark datasets. For example, the camera angle can be unusual, and patients can be partially covered with bedding sheets and electrode sets. Being able to track a patient in real-time with video EEG would be a promising innovation towards improving the quality of healthcare. Specifically, an automated patient detection system could supplement clinical oversight and reduce the resource-intensive efforts of nurses and doctors who need to continuously monitor patients. We evaluate an ImageNet pre-trained Mask R-CNN, a standard deep learning model for object detection, on the task of patient detection using our own curated dataset of 45 videos of hospital patients. The dataset was aggregated and curated for this work. We show that without fine-tuning, ImageNet pre-trained Mask R-CNN models perform poorly on such data. By fine-tuning the models with a subset of our dataset, we observe a substantial improvement in patient detection performance, with a mean average precision of 0.64. We show that the results vary substantially depending on the video clip.

Florian Dubost, Erin Hong, Max Pike et al.

Optimization plays a key role in the training of deep neural networks. Deciding when to stop training can have a substantial impact on the performance of the network during inference. Under certain conditions, the generalization error can display a double descent pattern during training: the learning curve is non-monotonic and seemingly diverges before converging again after additional epochs. This optimization pattern can lead to early stopping procedures to stop training before the second convergence and consequently select a suboptimal set of parameters for the network, with worse performance during inference. In this work, in addition to confirming that double descent occurs with small datasets and noisy labels as evidenced by others, we show that noisy labels must be present both in the training and generalization sets to observe a double descent pattern. We also show that the learning rate has an influence on double descent, and study how different optimizers and optimizer parameters influence the apparition of double descent. Finally, we show that increasing the learning rate can create an aliasing effect that masks the double descent pattern without suppressing it. We study this phenomenon through extensive experiments on variants of CIFAR-10 and show that they translate to a real world application: the forecast of seizure events in epileptic patients from continuous electroencephalographic recordings.

Siyi Tang, Jared A. Dunnmon, Khaled Saab et al.

Automated seizure detection and classification from electroencephalography (EEG) can greatly improve seizure diagnosis and treatment. However, several modeling challenges remain unaddressed in prior automated seizure detection and classification studies: (1) representing non-Euclidean data structure in EEGs, (2) accurately classifying rare seizure types, and (3) lacking a quantitative interpretability approach to measure model ability to localize seizures. In this study, we address these challenges by (1) representing the spatiotemporal dependencies in EEGs using a graph neural network (GNN) and proposing two EEG graph structures that capture the electrode geometry or dynamic brain connectivity, (2) proposing a self-supervised pre-training method that predicts preprocessed signals for the next time period to further improve model performance, particularly on rare seizure types, and (3) proposing a quantitative model interpretability approach to assess a model's ability to localize seizures within EEGs. When evaluating our approach on seizure detection and classification on a large public dataset, we find that our GNN with self-supervised pre-training achieves 0.875 Area Under the Receiver Operating Characteristic Curve on seizure detection and 0.749 weighted F1-score on seizure classification, outperforming previous methods for both seizure detection and classification. Moreover, our self-supervised pre-training strategy significantly improves classification of rare seizure types. Furthermore, quantitative interpretability analysis shows that our GNN with self-supervised pre-training precisely localizes 25.4% focal seizures, a 21.9 point improvement over existing CNNs. Finally, by superimposing the identified seizure locations on both raw EEG signals and EEG graphs, our approach could provide clinicians with an intuitive visualization of localized seizure regions.

Florian Dubost, Erin Hong, Nandita Bhaskhar et al.

Labeled data is a critical resource for training and evaluating machine learning models. However, many real-life datasets are only partially labeled. We propose a semi-supervised machine learning training strategy to improve event detection performance on sequential data, such as video recordings, when only sparse labels are available, such as event start times without their corresponding end times. Our method uses noisy guesses of the events' end times to train event detection models. Depending on how conservative these guesses are, mislabeled samples may be introduced into the training set. We further propose a mathematical model for explaining and estimating the evolution of the classification performance for increasingly noisier end time estimates. We show that neural networks can improve their detection performance by leveraging more training data with less conservative approximations despite the higher proportion of incorrect labels. We adapt sequential versions of CIFAR-10 and MNIST, and use the Berkeley MHAD and HMBD51 video datasets to empirically evaluate our method, and find that our risk-tolerant strategy outperforms conservative estimates by 3.5 points of mean average precision for CIFAR, 30 points for MNIST, 3 points for MHAD, and 14 points for HMBD51. Then, we leverage the proposed training strategy to tackle a real-life application: processing continuous video recordings of epilepsy patients, and show that our method outperforms baseline labeling methods by 17 points of average precision, and reaches a classification performance similar to that of fully supervised models. We share part of the code for this article.

Jared Dunnmon, Alexander Ratner, Nishith Khandwala et al.

Labeling training datasets has become a key barrier to building medical machine learning models. One strategy is to generate training labels programmatically, for example by applying natural language processing pipelines to text reports associated with imaging studies. We propose cross-modal data programming, which generalizes this intuitive strategy in a theoretically-grounded way that enables simpler, clinician-driven input, reduces required labeling time, and improves with additional unlabeled data. In this approach, clinicians generate training labels for models defined over a target modality (e.g. images or time series) by writing rules over an auxiliary modality (e.g. text reports). The resulting technical challenge consists of estimating the accuracies and correlations of these rules; we extend a recent unsupervised generative modeling technique to handle this cross-modal setting in a provably consistent way. Across four applications in radiography, computed tomography, and electroencephalography, and using only several hours of clinician time, our approach matches or exceeds the efficacy of physician-months of hand-labeling with statistical significance, demonstrating a fundamentally faster and more flexible way of building machine learning models in medicine.