PaperlayerView exact revision on arXiv

Source-linked AI summary

DeepCare: A Deep Dynamic Memory Model for Predictive Medicine

Trang Pham, Truyen Tran, Dinh Phung, Svetha Venkatesh

arXiv:1602.00357v2stat.MLcs.LG

TL;DR

Predictive medicine needs models that capture long-term, irregular healthcare trajectories and the influence of interventions. DeepCare combines learned admission representations, time-aware LSTM memory, intervention modeling, and multiscale pooling, improving results across diabetes and mental-health prediction tasks while requiring broader evaluation across cohorts and sites.

  • Problem

    Predictive healthcare must model long-term illness and care dependencies in episodic, irregular EMR observations, but existing methods inadequately capture this complexity.

  • Method

    DeepCare is an end-to-end neural network that learns admission vectors, parameterizes time in LSTM memory, incorporates interventions, and pools illness states across temporal scales.

  • Results

    DeepCare outperforms state-of-the-art baseline classification methods across disease progression, intervention recommendation, and future risk prediction on diabetes and mental health cohorts.

  • Takeaways & Limitations

    DeepCare provides a single EMR-based framework for memorizing illness trajectories, inferring present illness states, and predicting future medical outcomes without manual feature extraction.

  • Takeaways & Limitations

    More extensive evaluations across varied cohorts, sites, and outcomes are necessary.

Abstract

from arXiv · show

Personalized predictive medicine necessitates the modeling of patient illness and care processes, which inherently have long-term temporal dependencies. Healthcare observations, recorded in electronic medical records, are episodic and irregular in time. We introduce DeepCare, an end-to-end deep dynamic neural network that reads medical records, stores previous illness history, infers current illness states and predicts future medical outcomes. At the data level, DeepCare represents care episodes as vectors in space, models patient health state trajectories through explicit memory of historical records. Built on Long Short-Term Memory (LSTM), DeepCare introduces time parameterizations to handle irregular timed events by moderating the forgetting and consolidation of memory cells. DeepCare also incorporates medical interventions that change the course of illness and shape future medical risk. Moving up to the health state level, historical and present health states are then aggregated through multiscale temporal pooling, before passing through a neural network that estimates future outcomes. We demonstrate the efficacy of DeepCare for disease progression modeling, intervention recommendation, and future risk prediction. On two important cohorts with heavy social and economic burden -- diabetes and mental health -- the results show improved modeling and risk prediction accuracy.

1 Introduction

DeepCare addresses long-term, irregular, intervention-dependent healthcare prediction by combining learned admission representations with time-aware recurrent memory and temporal pooling. It is evaluated on disease progression, intervention recommendation, and future risk prediction across diabetes and mental health cohorts.

  • Motivation: Healthcare prognosis requires modeling long-term dependencies, variable-size admission episodes, irregular recording times, and interactions between disease progression and interventions.Historical illness and care can influence future risks and available interventions, while hospital records are episodic and irregularly timed.
  • Motivation: Existing methods inadequately capture variable-length records and long-term dependencies, while Markovian models lack memory and cannot represent temporal irregularity.The paper motivates an end-to-end deep dynamic memory model to address these limitations.
  • Method: DeepCare embeds diagnoses and interventions into continuous vectors, pools elements by type, concatenates them into admission representations, and learns features without manual engineering.These admission vectors become input features for the LSTM.
  • Method: The model handles irregular timing by making forgetting depend on time gaps, using monotonic decay and full time-parameterization mechanisms.Only observed records are incorporated, regardless of irregular spacing.
  • Method: DeepCare aggregates LSTM illness states with time-decayed multiscale pooling before a neural network estimates future prognosis.Its end-to-end pipeline reads medical records, memorizes illness trajectories, infers current states, and predicts future risk.
  • Evaluation: The experiments cover disease progression, intervention recommendation, and future risk prediction on diabetes and mental health cohorts, with DeepCare outperforming state-of-the-art baseline classification methods.The cohorts were collected from a large regional hospital between 2002 and 2013.

2 Background

The paper presents EMRs as detailed, precisely timed records for predictive modeling, while emphasizing that healthcare trajectories are episodic, irregular, and interleaved with care processes. It situates DeepCare within deep learning approaches that aim to learn features automatically from complex medical data.

  • Electronic medical records: An EMR is a digital version of patients’ health information containing symptoms, monitoring data, and clinicians’ observations.EMR systems are described as storing health information accurately while reducing replication and data-loss risks.
  • Electronic medical records: Typical EMRs record sequences of admissions with admission and discharge times, lab tests, diagnoses, procedures, medications, and clinical narratives.Diagnoses, procedures, and medications are typically represented using standardized codes.
  • Deep learning for healthcare: DeepCare addresses manual feature engineering by learning features automatically from EMR data in an end-to-end system.This positions the model as a response to the complexity of healthcare records.
  • Healthcare trajectories: Illness trajectories are important to model, but qualitative trajectory models are imprecise in time and therefore have limited predictive power.The paper contrasts this with the precise event timing available in EMRs.
  • Healthcare trajectories: EMRs provide a quantitative alternative for illness trajectories, but their records are complex because illness and care processes interleave and observations arrive irregularly.Patients may visit hospitals after illness passes a threshold even though the illness began earlier.
  • Existing modeling approaches: Markov models and dynamic Bayesian networks have been widely used for disease progression, but healthcare is non-Markovian because long-term dependencies, irregular timing, and interventions are not adequately modeled.A routine irrelevant admission can destroy the effect of a severe prior illness in Markov-style modeling.
  • Deep learning for healthcare: Deep learning has achieved major successes in speech, vision, and natural language processing, while its healthcare promise remained largely unrealized in the cited background.The paper notes that prior healthcare work did not properly model irregular timing.

3 Preliminaries

The preliminaries introduce recurrent neural networks and LSTMs for modeling sequential data and long-term dependencies. They also describe embeddings and pooling strategies for converting discrete inputs and variable-length sequences into representations for prediction.

  • 3.1 Recurrent neural network: RNNs process sequences through recurrent connections, shared parameters, and hidden states that summarize previous inputs.Outputs can be transformed into class probabilities using logistic or softmax functions.
  • 3.1 Recurrent neural network: RNNs are difficult to train on long dependencies because gradients can vanish or explode during back-propagation through time.LSTM is introduced as an effective approach for handling very long sequences.
  • 3.2 Long-short term memory: LSTM adds a memory cell and input, forget, and output gates to control information flow through the sequence.The forget gate determines how much past memory is retained, while the output gate controls information emitted from the cell.
  • 3.2 Long-short term memory: The additive LSTM memory pathway supports long-term memorization by reducing vanishing or exploding gradients.When the forget gate approaches 1, past memory is preserved; when it approaches 0, the system becomes memoryless.
  • LSTM for sequence classification: Mean pooling over LSTM output states can support sequence classification with a differentiable loss function.For binary outcomes, the pooled representation can be passed to logistic regression.
  • 3.3 Learning word representation: Word embeddings map discrete elements into lower-dimensional continuous vectors, while pooling aggregates vectors into fixed-size representations.The paper uses these ideas as groundwork for representing diagnoses and other discrete medical-record elements.

4 DeepCare

DeepCare extends LSTM into an end-to-end model for medical records with variable-size admissions, interventions, and irregular timing. It embeds admission codes, models admission context in the recurrent layer, pools illness states across scales, and estimates outcome probabilities.

  • 4 DeepCare: DeepCare extends LSTM to address variable-size discrete inputs, intervention effects, and irregular timing in illness trajectories.The model is designed for modeling illness trajectories and predicting future outcomes.
  • Architecture: Each admission is represented by diagnosis and intervention vectors together with its admission method and elapsed time.The input sequence uses u_t = [x_t, p_t, m_t, Δ_t].
  • DeepCare architecture: DeepCare aggregates illness states through multiscale weighted pooling before a neural network estimates the final outcome probability.The output probability can use logistic, softmax, or Gaussian structures for binary, multiclass, or continuous outcomes.
  • Representing variable-size admissions: Diagnosis and intervention codes are embedded separately and pooled into a 2M-dimensional admission vector [x_t, p_t].Pooling may use max, sum, or mean operations; normalized sum pooling reduces sensitivity to variable admission length.
  • Representing variable-size admissions: The admission method modifies the input gate so unplanned admissions can influence how much new information enters memory.Unplanned admissions are described as emergency transfers and typically indicate higher risk.

4.3 Modeling effect of interventions

DeepCare models interventions as factors that alter both current illness-state output and subsequent memory forgetting. This represents interventions as influences with potentially immediate and long-term effects.

  • 4.3 Modeling effect of interventions: The intervention vector moderates the output gate, changing the illness state produced at the current admission.The intervention weight matrix for this gate is P_o.
  • 4.3 Modeling effect of interventions: Previous interventions moderate the forget gate to represent long-term effects such as disease treatment or toxicity.The previous intervention embedding p_{t−1} is combined with an intervention weight matrix P_f.

4.4 Capturing time irregularity

DeepCare adapts memory forgetting to irregular intervals between admissions. It offers both monotonic time decay and a more flexible parametric forgetting mechanism for differing illness dynamics.

  • 4.4 Capturing time irregularity: The model modifies the forget gate so illness memory can change as time passes between admissions.This addresses the assumption that illness states remain constant over time.
  • 4.4 Capturing time irregularity: Monotonic decay reduces the forget gate as the elapsed time increases, modeling acute conditions whose effects naturally diminish.One tested decay function is d(Δ_{t−1:t}) = [log(e + Δ_{t−1:t})]^-1, with time measured in days.
  • 4.4 Capturing time irregularity: Parametric forgetting provides a more flexible time-dependent mechanism for chronic or worsening conditions.The time difference is transformed into a vector and weighted parametrically rather than imposing monotonic decay.

4.5 Prognosis through multiscale pooling and recency attention

DeepCare aggregates illness states across multiple look-back windows while weighting recent events more heavily, then uses the resulting representation to infer future prognosis.

  • Recency attention: Mean-pooling aggregates illness states but does not reflect healthcare’s emphasis on recency.The pooled state is computed from the sequence of illness states.
  • Recency attention: DeepCare introduces attention that assigns greater weight to recent events than older ones.The attention scheme uses elapsed time between each step and the current step.
  • Multiscale pooling: The pooling process supports one-year and other look-back windows by controlling the starting time step.A one-year look-back uses elapsed times of no more than 12 months.
  • Multiscale pooling: DeepCare uses 12-month, 24-month, and all-available-history windows to accommodate different disease progression rates.The three pooled illness states are stacked into one vector.
  • Prognosis inference: The stacked pooled illness-state vector is fed to a neural network that infers future prognosis.

4.6 Model complexity

DeepCare’s parameter count is determined by its vocabulary, embedding, recurrent, and neural-network dimensions, with additional time-related forget-gate parameters in the parametric-time variant.

  • Parameter count: DeepCare has M × |V| + M × K + K × K + K × D model parameters.The count comprises the model’s listed embedding, recurrent, and neural-network components.
  • Embedding parameters: Admission embedding uses two matrices, A and B, whose combined dimensions are M × |V|.
  • Gate parameters: The output gate contains input, recurrent, peephole, and bias parameters with dimensions M × K, K × K, K × K, and K × 1.
  • Gate parameters: The forget gate has analogous parameters, while parametric time adds a Qf matrix with Ntime = 3.Time decay adds no other parameters.
  • Neural-network parameters: The neural-network layer uses a 3K × D input-hidden matrix, a D × 2 hidden-output matrix, and two bias vectors.

4.7 Learning

After pooled illness states are formed, DeepCare applies a one-hidden-layer neural network to estimate outcomes and trains the resulting differentiable model by cross-entropy and back-propagation.

  • Forward model: The stacked illness-state vector is passed to a neural network with one hidden layer.
  • Objective: For binary classification, DeepCare models P(y | u0:n) with logistic regression and minimizes cross-entropy.The labels satisfy y ∈ {0, 1}.
  • Optimization: DeepCare’s loss is fully differentiable, enabling standard back-propagation with learning complexity linear in the number of parameters.
  • Forward pass: The forward pass reads patients’ disease-history records as its inputs.
  • Forward pass: At each time step, DeepCare embeds diagnoses and interventions, computes gates and illness states, and updates its recurrent representation.
  • Forward pass: The model then computes the pooled state, outcome probability, and loss function.

4.8 Pretraining and regularization

DeepCare pretrains risk-prediction representations with auxiliary diagnosis and intervention tasks, then uses L2 regularization and carefully placed dropout to reduce overfitting.

  • Pretraining: Auxiliary pretraining tasks predict next-readmission diagnoses and current-admission interventions for future risk prediction.These tasks also support disease-progression tracking.
  • Pretraining: The bottom LSTM layer produces illness states used to generate auxiliary diagnosis or intervention labels before initializing the risk-prediction embedding matrix.
  • Regularization: DeepCare uses L2-norm regularization and dropout because its intervention- and time-aware gates introduce three additional parameter matrices.
  • Regularization: Because standard dropout can damage sequential dependencies in RNNs, DeepCare applies dropout only at the input and neural-network layers.
  • Dropout placement: Dropout is applied to diagnosis and intervention codes before pooling, input features after admission representation, and neural-network input and hidden units.

5 Experiments

DeepCare is evaluated for disease progression, intervention recommendation, and future-risk prediction across diabetes and mental-health cohorts from a 12-year hospital dataset. It improves over memoryless, non-temporal, and simpler recurrent baselines across these tasks.

  • 5.1 Data: The study uses diabetes and mental-health cohorts collected over 12 years from a large Australian hospital.After preprocessing, the cohorts contained 7,191 diabetes patients with 53,208 admissions and 6,109 mental-health patients with 52,049 admissions.
  • 5.3 Disease progression: DeepCare is compared with Markov models and plain RNNs for next-diagnosis prediction using Precision@K.The evaluation predicts the next diagnoses at discharge and measures the percentage of relevant retrieved diagnoses.
  • 5.3 Disease progression: 2% improvement in diabetes and mental-health Precision@1 is obtained by DeepCare over plain RNNs for next-diagnosis prediction.For diabetes, plain RNN improves over Markov models by 8.8% at np = 1 and 27.7% at npred = 3; for mental health, Markov precision is 9.5% at np = 1 and RNN gains 50%.
  • 5.4 Intervention recommendation: DeepCare with sum-pooling outperforms the other models for current intervention prediction in both cohorts.Plain RNN consistently outperforms Markov models across values of np before DeepCare with sum-pooling achieves the best results.
  • 5.5 Future risk prediction: 79.0% F-score is achieved for diabetes unplanned-readmission prediction with parametric time, improving 7.6% over the best baseline.The best baseline is Random Forest with sum pooling at 71.4%; the mental-health best model improves 6.8% over a 67.9% sum-pooling Random Forest baseline.
  • 5.5 Future risk prediction: Nearly 60% F-score is reached for diabetes high-risk prediction with max-pooling DeepCare, while sum-pooling reaches 50.0% in mental health.RNN improves the best non-temporal model by over 10% F-score in both cohorts.

6 Discussion and Conclusion

DeepCare uses memory-inspired representations and recency mechanisms to model illness experiences across time, supporting disease progression, intervention recommendation, and prognosis from electronic medical records. The authors report competitive results across diabetes and mental-health tasks, while noting that broader evaluation remains necessary.

  • Discussion: Figure 8 evaluates high-risk prediction with F-score within 12 months for diabetes and 3 months for mental health, comparing DeepCare with mean, sum, and max admission pooling.The figure defines DC as DeepCare and the three pooling methods as admission-pooling baselines.
  • Discussion: The model’s embedding approach represents the semantics of diagnoses, interventions, and admissions using local contexts such as the next admission.This differs from prior embedding work that uses global contexts and can estimate embeddings independently of the task.
  • Discussion: DeepCare stores, updates, forgets, and manipulates illness experiences over time-stamped episodes before pooling them to infer current states and future prognosis.Its recency effect is implemented through forgetting and multiscale pooling with weights that decay over time.
  • Discussion: The authors identify the need for more extensive evaluations across varied cohorts, sites, and outcomes before broader implementation claims can be assessed.They also propose domain adaptation through parameter sharing and extensions for timed outcome sequences and more flexible time parameterizations.
  • Conclusion: DeepCare is presented as an end-to-end deep dynamic memory neural network that reads medical records, memorizes illness trajectories, estimates present illness states, and predicts future risk.The framework is intended to model disease progression, recommend interventions, and provide prognosis from electronic medical records.
  • Conclusion: The authors state that DeepCare achieves competitive results against current state-of-the-art methods on disease-stage prediction, intervention recommendation, and unplanned-readmission estimation in diabetic and mental-health patients.The paper describes these evaluations as demonstrations of its predictive-medicine framework.
Loading 1602.00357v2…