Analyzing the role of model uncertainty for electronic health records

Abstract

In medicine, both ethical and monetary costs of incorrect predictions can be significant, and the complexity of the problems often necessitates increasingly complex models. Recent work has shown that changing just the random seed is enough for otherwise well-tuned deep neural networks to vary in their individual predicted probabilities. In light of this, we investigate the role of model uncertainty methods in the medical domain. Using recurrent neural network (RNN) ensembles and various Bayesian RNNs, we show that population-level metrics, such as AUC-PR, AUC-ROC, log-likelihood, and calibration error, do not capture model uncertainty. Meanwhile, the presence of significant variability in patient-specific predictions and optimal decisions motivates the need for capturing model uncertainty. Understanding the uncertainty for individual patients is an area with clear clinical impact, such as determining when a model decision is likely to be brittle. We further show that RNNs with only Bayesian embeddings can be a more efficient way to capture model uncertainty compared to ensembles, and we analyze how model uncertainty is impacted across individual input features and patient subgroups. * Work done as a Google AI Resident.

On predictions in critical care: The individual prognostication fallacy in elderly patients

“…Consequently, irreversible decisions should not be solely based on group statistics. By using quantitative measures of uncertainty in individual cases [ 22 ], statistical predictions can still be considered for decision-making though, but in a more cautious and transparent way to adhere to the principles of medical ethics, especially non-maleficence and justice [ 21 ]. In addition to the traditional methods for prognostication, TLTs could provide a framework for individualised predictions in the future that is expected to reduce but not abolish predictive uncertainty.…”

“…However, this approach is based on group statistics and, thereby, affected by some fundamental limitations when applied to individual patients [ 11 ]. Of note, this problem also applies to new technologies from the field of artificial intelligence and machine learning [ 21 , 22 ]. First, it is assumed that an individual's characteristics match those of a chosen reference group that, importantly, is not uniform in itself.…”

Second opinion needed: communicating uncertainty in medical machine learning

“…Being able to estimate this posterior well should allow for good uncertainty estimates based on theoretical and empirical evidence 24,35 . Variational inference methods 36,37 are one popular class of approximations, but impose stricter assumptions about correlations between model parameters than more flexible methods 4,[38][39][40][41][42] . However, variational inference is known to underestimate the posterior probability distribution 43 .…”

“…There has been enormous progress towards the goal of medical artificial intelligence (AI) through the use of machine learning, resulting in a new set of capabilities on a wide variety of medical applications [1][2][3] . As these advancements translate into real-world clinical decision tools, many are taking stock of what capabilities these systems presently lack 4 , especially in light of some mixed results from prospective validation efforts 3,5,6 . While there are many possibilities, this article advocates that uncertainty quantification should be near the top of this list.…”

A prospective evaluation of AI-augmented epidemiology to forecast COVID-19 in the USA and Japan

“…For model reliability when used by human experts, we also investigate using epistemic model uncertainty as a confidence metric of forecasts. Well-calibrated estimates of uncertainty are important for being able to make more reliable predictions [102][103][104][105][106] . To this extent, we investigate the relationship between epistemic model uncertainty and the accuracy of forecasts by simulating the scenario of deciding whether or not to withhold each day's 28-day forecast based on model disagreement.…”