Multitask learning and benchmarking with clinical time series data

By Hrayr Harutyunyan, Hrant Khachatrian, David C. Kale, Aram Galstyan. 2017

1,081 citations recorded by the source corpus.

Health care is one of the most exciting frontiers in data mining and machine learning. Successful adoption of electronic health records (EHRs) created an explosion in digital clinical data available for analysis, but progress in machine learning for healthcare research has been difficult to measure because of the absence of publicly available benchmark data sets. To address this problem, we propose four clinical prediction benchmarks using data derived from the publicly available Medical Information Mart for Intensive Care (MIMIC-III) database. These tasks cover a range of clinical problems including modeling risk of mortality, forecasting length of stay, detecting physiologic decline, and phenotype classification. We propose strong linear and neural baselines for all four tasks and evaluate the effect of deep supervision, multitask training and data-specific architectural modifications on the performance of neural models.

Subfields: Artificial Intelligence, Epidemiology, Health Information Management

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Epidemiology

In Health Information Management

Citations in the corpus

A Communication Efficient Vertical Federated Learning Framework (2019, positive): “Following the data processing procedures of [17],”
Interpretable and Steerable Sequence Learning via Prototypes (2019, methodology): “They have been successfully applied to sentiment analysis [34], ECG signal classification [16], mortality and disease risk prediction using EHR data [8, 10, 13], and etc.”
Analyzing the role of model uncertainty for electronic health records (2019, background): “Existing state-of-the-art approaches using deep neural networks in medicine often make use of either a single model or an average over a small ensemble of models, focusing on improving the accuracy of probabilistic predictions [4, 14, 26, 35].”
Scalable and accurate deep learning with electronic health records (2018, neutral): “tured and semi-structured data. Because of the availability of Medical Information Mart for Intensive Care (MIMIC) data [48], many prior studies also have focused on ICU patients from a single center [92, 41]; other single-center studies have also focused on ICU patients [61]. Each ICU patient has s”
A Review of Challenges and Opportunities in Machine Learning for Health. (2018, background): “For example, inferring drug-resistance status in tuberculosis from whole-genome sequencing data 103 , predicting cardiovascular risk from EHR data 64,104 , readmission and mortality risk from longitudinal EHR data 39,105 , and predicting cardiovascular risk from retinal images 106 .”
Hurtful words: quantifying biases in clinical contextual word embeddings (2020, methodology): “For all tasks, we use the same held-out test set defined in previous work [27].”
Identify Susceptible Locations in Medical Records via Adversarial Attacks on Deep Predictive Models (2018, background): “In previous works, this problem was addressed by taking hourly averages of each feature over the course of an admission [17].”
Multi-modal Deep Learning (2024, methodology): “To address this issue, the MIMIC-III benchmark from [17] provides a standardized data cleaning and extraction process for the MIMIC-III database.”
Self-Supervised Contrastive Pre-Training For Time Series via Time-Frequency Consistency (2022, background): “Time series plays important roles in many areas, including clinical diagnosis, traffic analysis, and climate science [1, 2, 3, 4, 5, 6].”
Disease-Atlas: Navigating Disease Trajectories using Deep Learning (2018, background): “This includes joint architectures for forecasting follow-up times and clinical diagnoses (Choi et al., 2016; Du et al., 2016) and multitask learning (Harutyunyan et al., 2017; Razavian et al., 2016; Lipton et al., 2016).”
Artificial neural networks condensation: A strategy to facilitate adaption of machine learning in medical settings by reducing computational burden (2018, methodology): “We excluded all neonatal and pediatric patients (age 18 or younger at time of ICU stay) because the physiology of pediatric critical care patients differs signiﬁcantly from adults [13].”
Attend and Diagnose: Clinical Time Series Analysis using Attention Models (2017, methodology): “Hence, similar to the approach in (Harutyunyan et al. 2017), we implemented a multi-task version of our approach, SAnD-Multi, that uses a loss function that jointly evaluates the performance of all tasks, which can be expressed as follows:”
SOM-VAE: Interpretable Discrete Representation Learning on Time Series (2019, background): “Each sample in the time-grid was then labeled using a dynamic variant of the APACHE score (Knaus et al., 1985), which is a proxy for the instantaneous physiological state of a patient in the ICU.”
SCEHR: Supervised Contrastive Learning for Clinical Risk Prediction using Electronic Health Records (2021, methodology): “(multi-label or binary) cross-entropy loss] [5]–[7].”
Updates in Human-AI Teams: Understanding and Addressing the Performance/Compatibility Tradeoff (2019, background): “2016), in-hospital mortality prediction (Will a patient die in the hospital?) (Johnson et al. 2016; Harutyunyan et al. 2017), and credit risk assessment (Will a borrower fail to pay back?)3.”
Benchmark of Deep Learning Models on Large Healthcare MIMIC Datasets (2017, methodology): “With the goal of accurately predicting these clinical outcomes, researchers have developed novel machine learning models [1, 2] and scoring systems [3] while measuring the”
Designing Theory-Driven User-Centric Explainable AI (2019, methodology): “Hence we used the MIMIC III [45] dataset to train a multi-label gradient boosted tree (XGBoost [17]) using 833 extracted features (17 patient vital × 7 time windows × 7 statistics) [36] to perform phenotype diagnosis based on the first 24 hours vitals of a patient in the ICU.”