Andrew Beam

Bhawesh Kumar, Anil Palepu, Rudraksh Tuwani et al. · deepmind

Self-supervised models trained with a contrastive loss such as CLIP have shown to be very powerful in zero-shot classification settings. However, to be used as a zero-shot classifier these models require the user to provide new captions over a fixed set of labels at test time. In many settings, it is hard or impossible to know if a new query caption is compatible with the source captions used to train the model. We address these limitations by framing the zero-shot classification task as an outlier detection problem and develop a conformal prediction procedure to assess when a given test caption may be reliably used. On a real-world medical example, we show that our proposed conformal procedure improves the reliability of CLIP-style models in the zero-shot classification setting, and we provide an empirical analysis of the factors that may affect its performance.

Joe B Hakim, Jeffery L Painter, Darmendra Ramcharran et al.

Large language models (LLMs) are useful tools with the capacity for performing specific types of knowledge work at an effective scale. However, LLM deployments in high-risk and safety-critical domains pose unique challenges, notably the issue of ``hallucination,'' where LLMs can generate fabricated information. This is particularly concerning in settings such as drug safety, where inaccuracies could lead to patient harm. To mitigate these risks, we have developed and demonstrated a proof of concept suite of guardrails specifically designed to mitigate certain types of hallucinations and errors for drug safety, and potentially applicable to other medical safety-critical contexts. These guardrails include mechanisms to detect anomalous documents to prevent the ingestion of inappropriate data, identify incorrect drug names or adverse event terms, and convey uncertainty in generated content. We integrated these guardrails with an LLM fine-tuned for a text-to-text task, which involves converting both structured and unstructured data within adverse event reports into natural language. This method was applied to translate individual case safety reports, demonstrating effective application in a pharmacovigilance processing task. Our guardrail framework offers a set of tools with broad applicability across various domains, ensuring LLMs can be safely used in high-risk situations by eliminating the occurrence of key errors, including the generation of incorrect pharmacovigilance-related terms, thus adhering to stringent regulatory and quality standards in medical safety-critical environments.

Yining Hua, Fenglin Liu, Kailai Yang et al.

Objectieve:This review aims to deliver a comprehensive analysis of Large Language Models (LLMs) utilization in mental health care, evaluating their effectiveness, identifying challenges, and exploring their potential for future application. Materials and Methods: A systematic search was performed across multiple databases including PubMed, Web of Science, Google Scholar, arXiv, medRxiv, and PsyArXiv in November 2023. The review includes all types of original research, regardless of peer-review status, published or disseminated between October 1, 2019, and December 2, 2023. Studies were included without language restrictions if they employed LLMs developed after T5 and directly investigated research questions within mental health care settings. Results: Out of an initial 313 articles, 34 were selected based on their relevance to LLMs applications in mental health care and the rigor of their reported outcomes. The review identified various LLMs applications in mental health care, including diagnostics, therapy, and enhancing patient engagement. Key challenges highlighted were related to data availability and reliability, the nuanced handling of mental states, and effective evaluation methods. While LLMs showed promise in improving accuracy and accessibility, significant gaps in clinical applicability and ethical considerations were noted. Conclusion: LLMs hold substantial promise for enhancing mental health care. For their full potential to be realized, emphasis must be placed on developing robust datasets, development and evaluation frameworks, ethical guidelines, and interdisciplinary collaborations to address current limitations.

David R. Bellamy, Bhawesh Kumar, Cindy Wang et al. · deepmind

In this work we introduce Labrador, a pre-trained Transformer model for laboratory data. Labrador and BERT were pre-trained on a corpus of 100 million lab test results from electronic health records (EHRs) and evaluated on various downstream outcome prediction tasks. Both models demonstrate mastery of the pre-training task but neither consistently outperform XGBoost on downstream supervised tasks. Our ablation studies reveal that transfer learning shows limited effectiveness for BERT and achieves marginal success with Labrador. We explore the reasons for the failure of transfer learning and suggest that the data generating process underlying each patient cannot be characterized sufficiently using labs alone, among other factors. We encourage future work to focus on joint modeling of multiple EHR data categories and to include tree-based baselines in their evaluations.

Sophia M Rein, Jing Li, Miguel Hernan et al.

The g-formula can be used to estimate causal effects of sustained treatment strategies using observational data under the identifying assumptions of consistency, positivity, and exchangeability. The non-iterative conditional expectation (NICE) estimator of the g-formula also requires correct estimation of the conditional distribution of the time-varying treatment, confounders, and outcome. Parametric models, which have been traditionally used for this purpose, are subject to model misspecification, which may result in biased causal estimates. Here, we propose a unified deep learning framework for the NICE g-formula estimator that uses multitask recurrent neural networks for estimation of the joint conditional distributions. Using simulated data, we evaluated our model's bias and compared it with that of the parametric g-formula estimator. We found lower bias in the estimates of the causal effect of sustained treatment strategies on a survival outcome when using the deep learning estimator compared with the parametric NICE estimator in settings with simple and complex temporal dependencies between covariates. These findings suggest that our Deep Learning g-formula estimator may be less sensitive to model misspecification than the classical parametric NICE estimator when estimating the causal effect of sustained treatment strategies from complex observational data.

Bhawesh Kumar, Charlie Lu, Gauri Gupta et al.

As large language models continue to be widely developed, robust uncertainty quantification techniques will become crucial for their safe deployment in high-stakes scenarios. In this work, we explore how conformal prediction can be used to provide uncertainty quantification in language models for the specific task of multiple-choice question-answering. We find that the uncertainty estimates from conformal prediction are tightly correlated with prediction accuracy. This observation can be useful for downstream applications such as selective classification and filtering out low-quality predictions. We also investigate the exchangeability assumption required by conformal prediction to out-of-subject questions, which may be a more realistic scenario for many practical applications. Our work contributes towards more trustworthy and reliable usage of large language models in safety-critical situations, where robust guarantees of error rate are required.

Allen Schmaltz, Andrew Beam

We present a novel end-to-end language model for joint retrieval and classification, unifying the strengths of bi- and cross- encoders into a single language model via a coarse-to-fine memory matching search procedure for learning and inference. Evaluated on the standard blind test set of the FEVER fact verification dataset, classification accuracy is significantly higher than approaches that only rely on the language model parameters as a knowledge base, and approaches some recent multi-model pipeline systems, using only a single BERT base model augmented with memory layers. We further demonstrate how coupled retrieval and classification can be leveraged to identify low confidence instances, and we extend exemplar auditing to this setting for analyzing and constraining the model. As a result, our approach yields a means of updating language model behavior through two distinct mechanisms: The retrieved information can be updated explicitly, and the model behavior can be modified via the exemplar database.

Benjamin Kompa, Jasper Snoek, Andrew Beam

Uncertainty quantification for complex deep learning models is increasingly important as these techniques see growing use in high-stakes, real-world settings. Currently, the quality of a model's uncertainty is evaluated using point-prediction metrics such as negative log-likelihood or the Brier score on heldout data. In this study, we provide the first large scale evaluation of the empirical frequentist coverage properties of well known uncertainty quantification techniques on a suite of regression and classification tasks. We find that, in general, some methods do achieve desirable coverage properties on in distribution samples, but that coverage is not maintained on out-of-distribution data. Our results demonstrate the failings of current uncertainty quantification techniques as dataset shift increases and establish coverage as an important metric in developing models for real-world applications.

Allen Schmaltz, Andrew Beam

Many practical applications of AI in medicine consist of semi-supervised discovery: The investigator aims to identify features of interest at a resolution more fine-grained than that of the available human labels. This is often the scenario faced in healthcare applications as coarse, high-level labels (e.g., billing codes) are often the only sources that are readily available. These challenges are compounded for modalities such as text, where the feature space is very high-dimensional, and often contains considerable amounts of noise. In this work, we generalize a recently proposed zero-shot sequence labeling method, "binary labeling via a convolutional decomposition", to the case where the available document-level human labels are themselves relatively high-dimensional. The approach yields classification with "introspection", relating the fine-grained features of an inference-time prediction to their nearest neighbors from the training set, under the model. The approach is effective, yet parsimonious, as demonstrated on a well-studied MIMIC-III multi-label classification task of electronic health record data, and is useful as a tool for organizing the analysis of neural model predictions and high-dimensional datasets. Our proposed approach yields both a competitively effective classification model and an interrogation mechanism to aid healthcare workers in understanding the salient features that drive the model's predictions.