PaperlayerView exact revision on arXiv

Source-linked AI summary

MetaPred: Meta-Learning for Clinical Risk Prediction with Limited Patient Electronic Health Records

Xi Sheryl Zhang, Fengyi Tang, Hiroko Dodge, Jiayu Zhou, Fei Wang

arXiv:1905.03218v1cs.LGstat.ML

TL;DR

Clinical risk prediction from EHR is important but challenged by sparse, irregular records and limited patient samples. MetaPred learns across related risk-prediction tasks and can be fine-tuned on target EHRs; it outperforms comparison methods across the evaluated tasks, particularly for low-resource MCI and Alzheimer prediction.

  • Problem

    Clinical risk prediction from EHR must contend with challenging records and limited patient samples, which are costly or sometimes impossible to expand.

    Source: §1 · p. 1

  • Method

    MetaPred meta-trains a model-agnostic predictor across related risk tasks using a simulated target and objective-level adaptation, with optional fine-tuning on target EHRs.

    Source: §1 · p. 2, §3.2 · p. 5

  • Results

    MetaPred outperforms TransLearn and MultiLearn on all evaluated tasks, with larger gaps for low-resource MCI and Alzheimer prediction.

    Source: §4.3 · p. 7

  • Takeaways & Limitations

    The evaluation supports MetaPred as a predictor for low-resource clinical risks across the tested cognition-related disorders and source-target combinations.

    Source: §1 · p. 2, §6 · p. 9

    Meta-learning is constrained when source and target task distributions are too diverse, a concern for heterogeneous patient records across diseases.

    Source: §3.2 · pp. 4–5

Abstract

from arXiv · show

In recent years, increasingly augmentation of health data, such as patient Electronic Health Records (EHR), are becoming readily available. This provides an unprecedented opportunity for knowledge discovery and data mining algorithms to dig insights from them, which can, later on, be helpful to the improvement of the quality of care delivery. Predictive modeling of clinical risk, including in-hospital mortality, hospital readmission, chronic disease onset, condition exacerbation, etc., from patient EHR, is one of the health data analytic problems that attract most of the interests. The reason is not only because the problem is important in clinical settings, but also there are challenges working with EHR such as sparsity, irregularity, temporality, etc. Different from applications in other domains such as computer vision and natural language processing, the labeled data samples in medicine (patients) are relatively limited, which creates lots of troubles for effective predictive model learning, especially for complicated models such as deep learning. In this paper, we propose MetaPred, a meta-learning for clinical risk prediction from longitudinal patient EHRs. In particular, in order to predict the target risk where there are limited data samples, we train a meta-learner from a set of related risk prediction tasks which learns how a good predictor is learned. The meta-learned can then be directly used in target risk prediction, and the limited available samples can be used for further fine-tuning the model performance. The effectiveness of MetaPred is tested on a real patient EHR repository from Oregon Health & Science University. We are able to demonstrate that with CNN and RNN as base predictors, MetaPred can achieve much better performance for predicting target risk with low resources comparing with the predictor trained on the limited samples available for this risk.

1 INTRODUCTION

Clinical risk prediction from EHR matters for timely care, but EHR data are difficult to analyze and patient samples are often scarce. MetaPred applies meta-learning across related risk-prediction tasks to support prediction when target-task data are limited.

1 INTRODUCTION

  • Clinical risk models can help decision makers identify risks early so they can act in time to provide better care.EHR analysis faces sequentiality, sparsity, noisiness, and irregularity.Source: §1 · p. 1
  • Patient samples for a specific disease and EHR corpus are limited, while collecting more can be expensive or impossible.The paper also notes that many deadly diseases have poorly understood biological mechanisms.Source: §1 · p. 1
  • Transfer learning has been used to improve medical prediction with limited samples, including fine-tuning an ImageNet-trained model on 130K skin images.That model achieved classification performance comparable to well-trained dermatologists.Source: §1 · p. 1
  • Prior pretrained models encode general rather than medical-specific knowledge, and their effectiveness across clinical scenarios is difficult to judge.The passage discusses transformer and BERT models fine-tuned for unstructured medical data.Source: §1 · p. 2
  • MetaPred trains a model-agnostic meta-learner across related prediction tasks, with a simulated target, sequential parameter updates, and objective-level adaptation.The learned model can be further fine-tuned with target-domain EHRs.Source: §1 · p. 2

2 PROBLEM SETUP

MetaPred addresses a low-resource target risk by learning to transfer from related disease domains. Its setup distinguishes source domains from a target domain and uses episodes involving sources and a simulated target before target fine-tuning and prediction.

2 PROBLEM SETUP

  • The setup uses related, higher-resource disease domains to learn transfer for a low-resource Alzheimer’s risk-prediction target.The passage names Mild Cognitive Impairment and dementia as examples of source domains.Source: §2 · p. 2, Figure 1 · p. 2
  • The formal setup defines K source disease domains and one target domain, with positive and negative patient EHR data in each.The goal is to learn the target-domain predictor using limited target samples and source-domain data.Source: §2 · p. 2
  • From a domain-adaptation perspective, the setup frames the problem as choosing and optimizing the form of the source-data collection used by the predictor.The predictor is represented as f = (D_S, X; Θ).Source: §2 · p. 2
  • Meta-train and meta-test split episodes by labels, with the testing data drawn from a simulated target that is itself a source domain.The passage names these procedures as meta-train and meta-test, respectively.Source: §2 · p. 2
  • The workflow samples episodes from source and simulated-target domains, learns predictor parameters episode by episode, fine-tunes on the genuine target, then predicts its risk.The figure description identifies learner and meta-learner gradient-update loops in the meta-training process.Source: §2 · pp. 2–3, Figure 2 · p. 3

3 THE METAPRED FRAMEWORK

MetaPred uses model-agnostic meta-learning to prepare predictors for low-resource target EHR risk prediction. Its framework combines adaptation on simulated target tasks with source-domain supervision.

3.1 Risk Prediction Models

  • EHR visits are represented as binary medical-event vectors, and CNN or LSTM sequence models map patient records to disease-risk predictions.The CNN embeds sparse visit records before temporal convolution and pooling; the LSTM replaces the convolutional layer while retaining the embedding and MLP structure.Source: §3.1 · pp. 3–4

3.2 MetaPred Architecture

  • MetaPred uses source-task learning to initialize parameters that can adapt to held-out samples from a simulated target domain.The inner loop updates parameters on source losses; the outer loop uses simulated-target loss to update the shared initialization through a meta-gradient.Source: §3.2 · p. 4
  • Because heterogeneous disease domains can hinder optimization-level transfer, MetaPred adds source-domain supervision to the objective to compensate for limited fast adaptation.The source-loss contribution is balanced against simulated-target loss by hyperparameter µ.Source: §3.2 · pp. 4–5
  • At meta-test time, the learned parameters are adapted to the genuine target disease using a few gradient steps.Testing episodes include source-domain and genuine-target samples, and the target test set is used for evaluation.Source: §3.2 · p. 5

4 EXPERIMENTS

Experiments on OHSU EHR evaluate MetaPred across cognitive-disease prediction tasks, data-resource levels, and source-domain choices. MetaPred outperforms transfer-learning and multitask baselines across tasks, while performance also varies with the source domains used.

4.1 Dataset

  • The OHSU research data warehouse contains over 2.5 million patients and 20 million encounters; five cognitive conditions with more than 1,000 cases form the study domains.Dementia and Amnesia are source domains, while MCI, Alzheimer’s disease, and Parkinson’s disease are target domains; controls have other cognitive disorders.Source: §4.1 · p. 5
  • The experiments use age-matched cases and controls, two-year observation windows, and six-month prediction windows, with diagnosis histories mapped to 1,016 ICD-9 group codes.Cases and controls are matched within five years of age; the 10,989 distinct ICD-9 codes are mapped to their first three digits.Source: §4.1 · p. 5

4.2 Experimental Setup

  • MetaPred is evaluated with CNN and LSTM predictors against traditional, neural, fine-tuned, transfer-learning, and multitask baselines using held-out target patients.The low-resource setup fine-tunes on 20% of labeled target patients, while evaluation uses a 20% test split, five random runs, and average performance.Source: §4.2 · p. 6

4.3 Performance Evaluation

  • MetaPred outperforms TransLearn and MultiLearn on all three tasks, with larger gaps for MCI and Alzheimer’s when labeled target data are low.With 20% labeled target samples, MetaPred AUC is 0.7876±.02 and 0.7682±.01, versus 0.6874 ± .01 and 0.6755 ± .03 for the corresponding single-task versions.Source: §4.3 · p. 7
  • Across source-domain combinations, MetaPred variants generally outperform TransLearn, and source choice substantially affects MCI and Alzheimer’s F1 scores.The largest MCI F1 gap across source-domain choices can approach 0.25; Amnesia is generally more beneficial as a source than Parkinson’s.Source: §4.3 · p. 8, Figure 5 · p. 8
  • The learned patient representations separate the five studied diseases reasonably well in a t-SNE visualization of held-out patients.The visualization uses representations from before the final MLP layer and holds out 518 cases from each domain.Source: §4.3 · p. 8, Figure 6 · p. 8

5 RELATED WORK

Meta-learning uses experience from related tasks to support learning on a target task, and includes structure-based and optimization-based approaches. The paper identifies limited exploration of meta-learning in healthcare and proposes MetaPred for clinical risk prediction with low-resource EHRs.

5 RELATED WORK

  • Meta-learning leverages experience from related tasks to solve a learning problem in a target task, and aims for fast learning from few examples.The passage distinguishes structure-based approaches using distances or embeddings from optimization-based approaches that train a gradient procedure.Source: §5 · p. 8
  • Meta-learning had rarely been explored in healthcare despite the resource-limited nature of many medical problems.Source: §5 · p. 8
  • MetaPred is proposed to address clinical risk prediction with low-resource EHRs.Source: §5 · p. 9

6 CONCLUSION

MetaPred combines deep clinical-risk prediction with model-agnostic meta-learning and objective-level adaptation for low-resource EHR tasks. Across real-world evaluations involving five cognitive diseases, it showed superior performance with limited patient records, including outperforming fully supervised deep neural networks on MCI and Alzheimer risk prediction.

6 CONCLUSION

  • MetaPred with limited patient EHRs showed superior performance, even beating fully supervised deep neural networks on challenging MCI and Alzheimer risk prediction tasks.The conclusion reports these results across evaluations involving five cognitive diseases.Source: §6 · p. 9
  • MetaPred uses deep predictive modeling and model-agnostic meta-learning to leverage labeled medical records from high-resource domains.Source: §6 · p. 9
  • Objective-level adaptation combines fast optimization-level adaptation with supervision from the high-resource domain.Source: §6 · p. 9
  • The evaluation used real-world EHR data across five cognitive diseases and varied source-target combinations; future work will explore records spanning more than five years.Source: §6 · p. 9

Related papers

Loading 1905.03218v1…