PaperlayerView exact revision on arXiv

Source-linked AI summary

Predicting Clinical Events by Combining Static and Dynamic Information Using Recurrent Neural Networks

Cristóbal Esteban, Oliver Staeck, Yinchong Yang, Volker Tresp

arXiv:1602.02685v2cs.LGcs.AIcs.NE

TL;DR

The paper asks how to predict rejection, transplant loss, or death from clinical records containing both static patient information and dynamic visit sequences. It develops RNN-based models for kidney-transplant patients and compares them with feedforward and logistic-regression models. GRU-based recurrence performs best for endpoint prediction, whereas a feedforward network performs best for next-event prediction.

  • Problem

    The paper addresses predicting kidney-transplant rejection, loss, or death within 6 or 12 months from clinical records combining static and dynamic information.

  • Method

    The study develops RNN models that combine static patient variables with dynamic clinical sequences and compares them with feedforward and logistic-regression models.

  • Results

    GRU-based RNNs achieve the best endpoint-prediction performance, while a Feedforward Neural Network outperforms the recurrent models for future medication and laboratory-result prediction.

  • Takeaways & Limitations

    RNNs are most effective for the endpoint task, whereas next-event prediction favors a feedforward model when complex long-term dependencies are not relevant.

Abstract

from arXiv · show

In clinical data sets we often find static information (e.g. patient gender, blood type, etc.) combined with sequences of data that are recorded during multiple hospital visits (e.g. medications prescribed, tests performed, etc.). Recurrent Neural Networks (RNNs) have proven to be very successful for modelling sequences of data in many areas of Machine Learning. In this work we present an approach based on RNNs, specifically designed for the clinical domain, that combines static and dynamic information in order to predict future events. We work with a database collected in the Charité Hospital in Berlin that contains complete information concerning patients that underwent a kidney transplantation. After the transplantation three main endpoints can occur: rejection of the kidney, loss of the kidney and death of the patient. Our goal is to predict, based on information recorded in the Electronic Health Record of each patient, whether any of those endpoints will occur within the next six or twelve months after each visit to the clinic. We compared different types of RNNs that we developed for this work, with a model based on a Feedforward Neural Network and a Logistic Regression model. We found that the RNN that we developed based on Gated Recurrent Units provides the best performance for this task. We also used the same models for a second task, i.e., next event prediction, and found that here the model based on a Feedforward Neural Network outperformed the other models. Our hypothesis is that long-term dependencies are not as relevant in this task.

I. INTRODUCTION

Clinical records combine long, wide, time-evolving data with static patient information, creating a need for models tailored to clinical event prediction. The paper develops recurrent approaches that integrate both information types and compares them with prior feedforward methods.

  • Digitized clinical records increasingly overwhelm human experts with large volumes of available information.
  • Medical datasets are becoming longer and wider, requiring models that capture complex relationships among many time-evolving variables.
  • Clinical event prediction resembles language modeling but must handle simultaneous events, multiple patient sequences, static variables, and mixed Boolean and numeric data.
  • The study uses a large Charité Hospital kidney-failure dataset containing patients who underwent or awaited renal transplantation.
  • The endpoint task predicts rejection, transplant loss, or death within 6 or 12 months after each clinic visit.
  • The proposed RNN model combines static patient background with dynamic visit data, addressing a central feature of clinical datasets.
  • Prior Temporal Latent Embeddings use static information and the previous n events in a feedforward network, while the paper targets RNN-based integration of static and sequential data.

III. KIDNEY TRANSPLANTATION ENDPOINTS

Kidney transplantation requires complex, long-term clinical management, while the paper focuses on predicting rejection, graft loss, or death from patients’ longitudinal and static information. Such predictions are framed as potential support for difficult clinical decision-making.

  • Kidney transplant recipients face severe complications, lifelong medication use, and long-term specialized care.
  • Their medical records span years to decades and contain numerous diagnoses, symptoms, results, medications, and laboratory values.
  • The volume and complexity of transplant data make clinical decision-making difficult during routine consultations.
  • Computerized prediction of graft failure, death, and other endpoints could help identify patients needing intensified medical care.
  • Earlier risk detection may allow timely identification of toxicities, interactions, infections, comorbidities, and other complications.
  • The three major post-transplant endpoints are kidney rejection, graft loss, and patient death.
  • The study predicts these endpoints 6 and 12 months after each clinic visit using medical history, medication and laboratory sequences, and static patient information.

IV. RECURRENT NEURAL NETWORKS FOR CLINICAL EVENT PREDICTION

RNNs update a hidden state from current inputs and prior history, enabling sequence models to retain clinically relevant past information. The paper presents this as useful for long-term dependencies and for prediction after variable amounts of patient history.

  • RNNs compute each hidden state by combining the current input with the previous hidden state.
  • This recurrent update allows RNNs to remember past events relevant to predicting future outcomes.
  • Unlike feedforward networks with a specified time window, RNNs can in principle retain useful events from throughout a patient’s history.
  • RNNs can begin predicting for a new patient after the first clinic visit, whereas fixed-window feedforward models require the chosen number of prior visits.
  • The RNN output applies a parameter matrix to the hidden state and then an element-wise differentiable function to produce predicted outputs.

A. Standard Recurrent Neural Network

Standard RNNs update a hidden state from the previous state and current input, allowing past events to influence predictions. However, vanishing gradients limit their ability to capture dependencies far back in clinical sequences.

  • Standard RNN update: The hidden state combines the previous hidden state with the current input, allowing the network to encode past events.W and U are model parameter matrices in this update.
  • Clinical motivation: Long-term memory can be useful when earlier clinical events inform future patient outcomes.
  • Vanishing gradients: Standard RNNs struggle with distant dependencies because gradients exponentially decay during backpropagation through time.The text reports empirical difficulty remembering events around 5–10 time steps in the past.
  • Gated alternatives: Gating mechanisms were developed to alleviate vanishing gradients by controlling when and how the hidden state is updated.
  • LSTM units: LSTM units regulate the combination of current input and previous hidden state through input, forget, and output gates.Compared with standard RNN updates, LSTMs use more parameters and are computationally more expensive.

C. Gated Recurrent Units

GRUs use gating to adaptively capture dependencies at different time scales. They are less complex than LSTMs and have outperformed standard RNNs and LSTMs on many data sets in prior comparisons.

  • GRU motivation: GRUs were introduced to let recurrent units adaptively capture dependencies at different time scales.
  • GRU mechanism: The GRU formulation uses reset and update gates to regulate recurrent-state computation.The supplied equations define the reset gate and candidate hidden state; the text identifies z and r as update and reset gates.
  • Comparison with LSTMs: GRUs are less complex than LSTM units and in many cases perform better experimentally.
  • Prior evidence: Across multiple data sets, GRU units outperformed standard RNNs and LSTM units in most cases.

D. Combining RNNs with Static Data

The proposed architecture processes static patient information and dynamic visit data separately, then combines their hidden representations to predict clinical targets. It uses latent input representations, recurrent updates, and a binary cross-entropy objective.

  • Architecture: The architecture processes static patient features with a feedforward network and dynamic visit data with an RNN.Static features include examples such as gender, blood type, and cause of kidney failure.
  • Architecture: The hidden states from the static and recurrent branches are concatenated before the output layer produces predictions.
  • Latent representations: Raw static and visit inputs are transformed into latent representations before hidden-state computation.The static vector describes patient-level information, while the visit vector contains information recorded at time t.
  • Recurrent component: The recurrent update can use a standard RNN, LSTM, or GRU function.
  • Objective: Training uses a Bernoulli-likelihood cost function, also called binary cross-entropy, over the training data.The true and predicted target variables are denoted by y and ŷ in the supplied prose.

A. Data Pre-processing and Experimental Setup

The study encodes clinical variables for visit-level prediction, evaluates models on unseen patients using repeated random splits, and emphasizes AUPRC for infrequent endpoints. Binary laboratory encoding outperformed imputation with scaling or normalization.

  • Data representation: The data contain binary endpoints and medications, while laboratory results are real-valued measurements.
  • Pre-processing: Missing laboratory measurements were handled by comparing mean or median imputation with scaling or normalization.
  • Pre-processing: Approximately 5% higher AUROC and 3% higher AUPRC were obtained by encoding each laboratory value as high, normal, or low instead of imputing and scaling or normalizing.Unmeasured values receive zeros for all three event indicators, removing the need for imputation.
  • Prediction target: Each visit produces targets for kidney rejection, kidney loss, and death at both 6- and 12-month horizons.
  • Data set: The cleaned data include 2,061 patients and 193,111 clinic visits, with endpoint density of 7.3%.The passage also reports that 38.4% of patients experienced at least one endpoint event.
  • Data set: Dynamic information contains 6,566 variables, while static information contains 342 patient-level features.
  • Experimental setup: Models were evaluated on unseen patients using 60% training, 20% validation, and 20% test splits.
  • Evaluation: For infrequent target events, AUPRC is treated as more informative than AUROC because high precision is harder to obtain.Results report mean AUPRC and AUROC with standard errors over five identical random splits.

B. Results

The recurrent models outperformed the other evaluated models on combined endpoint prediction, with GRU + static achieving the best reported scores. The experiments also examined repeated winning configurations and endpoint-specific performance.

  • Recurrent models outperformed the other models on both AUROC and AUPRC for combined endpoint prediction.
  • The most repeated winning configuration used 100 hidden units, rank size 50, dropout 0.1, learning rate 0.1, and Adagrad.
  • Endpoint-specific evaluation was reported for GRU + static using AUPRC and AUROC.
  • GRU + static performed best for predicting patient death and had its worst AUPRC for kidney loss within six months.

C. Additional experiments

Additional experiments evaluated the models for next-visit event prediction, including recurrent models alongside the Temporal Latent Embeddings model and static-only feedforward predictions. Temporal Latent Embeddings still achieved better scores than the other evaluated models.

  • The next-visit task predicted laboratory analyses, their results, and medication prescriptions observed at each patient’s subsequent clinic visit.
  • The experiments reproduced prior Temporal Latent Embeddings evaluations while adding the RNN models introduced in this work.
  • Temporal Latent Embeddings still provided better scores than the other models in next-visit prediction.
  • Static embeddings represented predictions from a feedforward neural network using only each patient’s static information.

VI. CONCLUSION

The paper develops recurrent models that combine static and dynamic clinical information for kidney-transplant endpoint prediction. GRU-based recurrence performs best for endpoint prediction, while feedforward models remain preferable for future medication and laboratory-result prediction.

  • Contribution: RNN algorithms combine static and dynamic clinical variables to predict endpoints in patients who underwent kidney transplantation.The approach is intended to support physicians with clinical decision-making information.
  • Endpoint prediction: The RNN with GRUs combined with a Feedforward Neural Network provides the best score for endpoint prediction.
  • Next-event prediction: For predicting future medications and laboratory results, RNNs do not outperform a Feedforward Neural Network.The authors hypothesize that complex long-term dependencies are less relevant for this task, favoring attention to recent events.
  • Input representation: Binary encoding of real-valued input variables performs better than normalization with imputation for missing data.
  • Future work: Future experiments aim to predict more complex medication targets, including doses and intake patterns.The current second experiment predicts prescribed medications without dose or intake-pattern information.
  • Future work: The study does not yet use all available kidney-failure patient information, including biopsies and their results.The authors plan to add further information sources to the models.
Loading 1602.02685v2…