Abstract
We present a personalized and reliable prediction model for healthcare, which can provide individually tailored medical services such as diagnosis, disease treatment, and prevention. Our proposed framework targets at making personalized and reliable predictions from time-series data, such as Electronic Health Records (EHR), by modeling two complementary components: i) a shared component that captures global trend across diverse patients and ii) a patient-specific component that models idiosyncratic variability for each patient. To this end, we propose a composite model of a deep neural network to learn complex global trends from the large number of patients, and Gaussian Processes (GP) to probabilistically model individual time-series given relatively small number of visits per patient. We evaluate our model on diverse and heterogeneous tasks from EHR datasets and show practical advantages over standard time-series deep models such as pure Recurrent Neural Network (RNN).
Introduction
In the realm of precision medicine, the "one size fits all" approach of traditional AI models often falls short. Most modern healthcare models, such as Recurrent Neural Networks (RNNs), are population-based, learning a single set of rules from thousands of patients. However, patients are inherently heterogeneous; factors like genetics, lifestyle, and environment mean that individuals with identical symptoms may follow vastly different clinical trajectories.
Standard deep learning models frequently fail to capture this "idiosyncratic variability." Conversely, Gaussian Processes (GPs) are excellent at modeling individual trends and providing probabilistic uncertainty estimates—crucial for medical decision-making. However, GPs are computationally expensive for large datasets and struggle to learn complex global patterns. The Deep Mixed Effect Model using Gaussian Processes (DME-GP) bridges this gap by combining the pattern-recognition power of Deep Learning with the personalized nature of GPs.
Methodology: The Mixed Effect Decomposition
The core of DME-GP is a Mixed Effect Model decomposition. A patient’s health function $f(x)$ is represented as the sum of two distinct parts:
- The Global Component ($g$): A deep neural network (RNN or MLP) that captures broad, complex trends shared across the entire population.
- The Individual Component ($l$): A patient-specific Gaussian Process that models how a specific individual deviates from the global average.
Scalability and Deep Kernel Learning
To handle the massive scale of Electronic Health Records (EHR), the authors treat each patient's "random effect" as independent. This transforms a computationally impossible global matrix into a block-diagonal one, allowing computation to scale linearly with the number of patients. Furthermore, the GP utilizes the hidden layers of the deep network as its input (Deep Kernel Learning), enabling it to process high-dimensional medical features effectively.
Datasets and Evaluation
The model was rigorously evaluated on two large-scale datasets:
- Vital-Sign Dataset (Physionet Challenge 2012): Heart rate time-series for 865 patients in cardiac surgery recovery.
- Medical Checkup Dataset (NHIS): Data from 32,927 patients in South Korea, used to predict the risk of 12 common diseases (e.g., Liver Cirrhosis, Heart Failure) based on historical checkups.
Key Insights and Results
DME-GP treats personalization as a "deviation" from a global baseline. Instead of requiring massive amounts of data per patient to train a unique network, it uses the GP to "nudge" the global prediction based on an individual's history.
Performance and Reliability
- Disease Prediction: DME-GP achieved an average AUC of 0.778 across 12 tasks, significantly outperforming standard RNNs (0.720) and other state-of-the-art models like RETAIN (0.715).
- Heart Rate Prediction: In regression tasks, DME-GP achieved a lower RMSE (0.150) compared to a standalone MLP (0.194) or a personalized GP without a global mean (0.243).
- Uncertainty as a Safety Feature: A key advantage of the GP component is its ability to produce confidence intervals. The model is naturally less confident on "hard" patients whose history is ambiguous, providing a vital signal for clinical intervention.
Conclusion and Future Work
While DME-GP scales linearly with the number of patients, it retains a cubic cost relative to the number of visits per patient. This is generally acceptable for EHR data but may pose challenges for high-frequency sensor data. Future research will explore more sophisticated deep architectures for the global component and apply the framework to multi-modal data, such as combining medical imaging with EHR.
Summary updated