Source-linked AI summary
Learning to Detect Sepsis with a Multitask Gaussian Process RNN Classifier
Joseph Futoma, Sanjay Hariharan, Katherine Heller
TL;DR
Early sepsis detection is difficult because clinical data are uncertain, frequently missing, irregularly sampled, and collected across highly variable encounters. The paper connects a multitask Gaussian process to an RNN and trains the system end-to-end, achieving improved prediction over NEWS and other baselines. Its results support using uncertainty-aware modeling of multivariate clinical time series for earlier sepsis prediction, while leaving heterogeneity, medication effects, covariance flexibility, and computational overhead as extension areas.
Problem
Early and accurate sepsis identification remains difficult despite routinely captured EHR data, because clinical time series are irregular, incomplete, heterogeneous, and have uncertain onset times.
Method
A multitask Gaussian process converts irregular, missing multivariate physiology into an uncertainty-preserving grid representation that feeds a recurrent neural network, with both trained discriminatively end-to-end.
Results
At sensitivity 0.85, the method predicts sepsis roughly 4 hours in advance with precision 0.64 and yields higher performance than NEWS and other baselines.
Takeaways & Limitations
The framework provides earlier sepsis prediction with higher precision and fewer false alarms than the evaluated clinical warning scores and modeling baselines.
Takeaways & Limitations
The shared correlation function across physiological streams may be overly restrictive, and the authors identify patient heterogeneity, medication effects, and computational overhead as areas for improvement.
Abstract
from arXiv · showhide
We present a scalable end-to-end classifier that uses streaming physiological and medication data to accurately predict the onset of sepsis, a life-threatening complication from infections that has high mortality and morbidity. Our proposed framework models the multivariate trajectories of continuous-valued physiological time series using multitask Gaussian processes, seamlessly accounting for the high uncertainty, frequent missingness, and irregular sampling rates typically associated with real clinical data. The Gaussian process is directly connected to a black-box classifier that predicts whether a patient will become septic, chosen in our case to be a recurrent neural network to account for the extreme variability in the length of patient encounters. We show how to scale the computations associated with the Gaussian process in a manner so that the entire system can be discriminatively trained end-to-end using backpropagation. In a large cohort of heterogeneous inpatient encounters at our university health system we find that it outperforms several baselines at predicting sepsis, and yields 19.4% and 55.5% improved areas under the Receiver Operating Characteristic and Precision Recall curves as compared to the NEWS score currently used by our hospital.
1. Introduction
The paper targets earlier, more accurate sepsis detection from heterogeneous EHR time series, addressing irregular sampling, missing values, uncertain onset, and variable encounters. It connects a multitask Gaussian process to an RNN and reports improved prediction over clinical and modeling baselines.
- 1. Introduction: Early sepsis identification matters because delayed treatment is associated with worsening outcomes, while accurate recognition remains difficult for clinicians.Sepsis symptoms may arise from many other clinical conditions.
- 1. Introduction: NEWS and related warning scores use limited information and independently score variables, overlooking relationships among physiology and its evolution over time.
- 1. Introduction: The proposed classifier transforms raw multivariate physiology through a Multitask Gaussian Process and feeds latent values to a recurrent neural network for binary sepsis prediction.The framework is designed for heterogeneous encounters with irregular sampling, missing values, uncertain sepsis onset, and variable clinical states.
- 1. Introduction: 4 hours in advance, the method predicts sepsis at sensitivity 0.85 and precision 0.64 on heterogeneous inpatient encounters from 18 months of Duke EHR data.
- 1. Introduction: 4.3% in area under the ROC curve and 11.1% in area under the Precision Recall curve are gained over an RNN trained on raw clinical data without a Gaussian Process.The paper also reports substantially better performance than commonly used early warning scores, including NEWS.
2. Related Works
Related work includes clinical warning scores, disease-progression models, multivariate time-series classifiers, recurrent neural networks, and multitask Gaussian processes. The paper differs by combining an MGP with end-to-end discriminative classification for heterogeneous, irregular, and incomplete clinical time series.
- 2. Related Works: MEWS, NEWS, SIRS, SOFA, qSOFA, and the Rothman Index are established scores for clinical deterioration or sepsis-related assessment.
- 2. Related Works: Prior EHR progression models address conditions including scleroderma, chronic kidney disease, cardiac events, and glaucoma, while a Cox model predicts sepsis without temporal structure.
- 2. Related Works: Earlier multivariate time-series classifiers often cluster series using ad-hoc distance metrics, whereas related clinical RNN studies use LSTM or GRU architectures.
- 2. Related Works: Multitask Gaussian processes have been used to forecast vitals and predict clinical instability, while this paper applies them to binary early sepsis classification.
- 2. Related Works: The paper builds on end-to-end discriminative learning of MGP and classifier parameters for multivariate, variable-length series with substantial missingness.
3. Proposed Model
The proposed model converts irregular, incomplete multivariate clinical time series into a standardized representation with a multitask Gaussian process, then uses an RNN to continuously predict sepsis risk. It propagates Gaussian-process uncertainty into an end-to-end classifier that incorporates physiology, admission covariates, and medications.
- Multitask Gaussian Processes: The multitask covariance structure allows information from frequently sampled physiological variables to improve learning for variables measured infrequently or not at all.The model separates covariance across variables from covariance across observation times and uses an Ornstein-Uhlenbeck temporal kernel.
- Multitask Gaussian Processes: The model receives irregularly sampled laboratory and vital-sign measurements and uses a multitask Gaussian process to produce latent values on shared, evenly spaced time grids.The Gaussian process maintains uncertainty while interpolating and imputing missing values, enabling a uniform representation across encounters.
- Classification Method: An LSTM-based recurrent neural network maps variable-length encounter sequences to sepsis probabilities using latent physiology, baseline covariates, and time-binned medication counts.This design captures time-varying interactions among admission characteristics, physiological measurements, and medication administration.
- Classification Method: Because the latent physiological values are uncertain, training minimizes expected classification loss under the multitask Gaussian-process posterior rather than treating imputed values as fixed observations.The resulting risk score is the posterior expectation of the RNN output and can be updated continuously as new encounter information arrives.
- End-to-End Learning Framework: Monte Carlo sampling and the reparameterization trick make gradients with respect to both recurrent-network and Gaussian-process parameters tractable for joint discriminative training.The model learns the Gaussian-process and classifier parameters together instead of fitting the Gaussian process first and fixing it before classifier training.
4. Experiments
Experiments evaluate MGP-RNN on heterogeneous EHR encounters against clinical scores and neural-network and Gaussian-process baselines. The framework generally performs best, particularly near sepsis onset, while precision is assessed at fixed sensitivity.
- Dataset: The dataset contains 49,312 inpatient admissions and 34 physiological variables extracted from 18 months of university health-system EHR data.The variables comprise 5 vital signs and 29 laboratory values with highly variable measurement frequency.
- Experimental setup: Sepsis encounters exclude data after acquisition, while non-septic encounters use data through discharge; risk scores refresh hourly from admission.The dataset is split into 80% training, 10% validation, and 10% testing encounters.
- Baselines: MGP-RNN is compared with NEWS, MEWS, SIRS, Raw RNN, MGP-RNN-mean, and GP-RNN baselines.The clinical scores use only small subsets of the available variables, while Raw RNN uses hourly imputation without a Gaussian Process.
- Evaluation: AU-ROC and AU-PR are evaluated across prediction horizons, while precision is evaluated across time at fixed sensitivity of 0.85.Figure 3 reports these metrics for predictions issued 0–12 hours before sepsis or discharge.
- Results: MGP-RNN generally performs best, followed by the other MGP/GP-RNN methods, with the largest separation from other methods in the final four hours.The results indicate that MGP-RNN retains uncertainty information and models correlations among physiological variables.
- Results: At sensitivity 0.85, MGP-RNN has its largest precision gains approximately 3–7 hours before sepsis and substantially higher precision than NEWS, MEWS, and SIRS near onset.The comparison highlights a precision–timeliness tradeoff relevant to reducing false alarms.
5. Conclusions and Clinical Significance
The paper presents an uncertainty-aware classifier for early sepsis detection and reports substantially better performance than clinical benchmarks. The authors connect this performance to higher precision and fewer false alarms, while identifying several extensions for heterogeneous data and computational efficiency.
- Clinical significance: At sensitivity 0.85 and four hours in advance, MGP-RNN produces roughly 0.5 false alarms per true alarm versus about 2.5 for NEWS.The authors state that this would substantially reduce total false alarms.
- Conclusions: The proposed approach classifies multivariate clinical time series while accounting for uncertainty in the series.It is evaluated on a large inpatient-encounter dataset from the authors’ university health system.
- Future directions: The method could be extended with encounter-subpopulation clustering, medication-effect modeling, richer covariance structures, and additional Gaussian-process approximations.The authors note that a shared correlation function across physiological streams may be overly restrictive and training is computationally expensive.
- Clinical significance: The paper argues that improved precision and lower false-alarm rates could reduce alarm fatigue and potentially improve patient outcomes.The clinical significance is motivated by the difficulty of early sepsis identification and the harms of delayed treatment.