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 era of precision medicine, the "one size fits all" approach is increasingly inadequate. Electronic Health Records (EHR) are inherently heterogeneous; patients presenting with identical symptoms may have vastly different biological profiles, environmental exposures, or disease trajectories.
Existing deep learning models, such as standard Recurrent Neural Networks (RNNs), are primarily population-based. They excel at identifying broad trends across thousands of patients but often fail to capture the individual nuances that define personalized care. Conversely, traditional statistical methods like Gaussian Processes (GPs) are highly effective at modeling individual variability and providing uncertainty estimates, but they suffer from cubic computational complexity ($O(n^3)$), making them unscalable for large-scale multi-task healthcare datasets. This work introduces a framework that bridges this gap, combining the pattern-recognition power of deep learning with the personalized precision of GPs.
Methodology
The proposed Deep Mixed Effect Model using Gaussian Processes (DME-GP) utilizes a functional decomposition of a patient's health trajectory:
$$f^{(i)}(x) = g(x) + l^{(i)}(x)$$
- Global Component ($g(x)$): A deep neural network (RNN or MLP) acts as a "fixed effect," learning complex, shared trends across the entire patient population.
- Individual Component ($l^{(i)}(x)$): A personalized Gaussian Process acts as a "random effect," capturing how a specific patient deviates from the global average.
To ensure scalability, the model treats each patient’s GP as independent conditioned on the shared global mean. This reduces complexity from $O((PT)^3)$ to $O(PT^3)$, where $P$ is the number of patients and $T$ is the number of visits. Furthermore, the hidden states of the deep model are used as embeddings for the GP kernel, creating a "Deep Kernel" that enhances the expressiveness of the local model.
Experimental Results
The model was evaluated on two distinct EHR datasets: the Physionet Challenge 2012 Vital-Sign dataset (regression) and a large-scale Medical Checkup dataset from NHIS (classification).
Heart Rate Prediction
In predicting heart rate progressions for cardiac surgery recovery patients, DME-GP achieved a Root Mean Square Error (RMSE) of 0.150, significantly outperforming standard MLPs (0.194) and standalone personalized GPs (0.243). As shown in the qualitative analysis, DME-GP tracks the true patient trajectory more accurately than population-based or purely local models.
Disease Risk Prediction
Across 12 common diseases, DME-GP consistently outperformed standard RNNs and meta-learning baselines. For Liver Cirrhosis prediction, it achieved an AUC of 0.932, compared to 0.888 for a standard RNN. A key finding was the model's reliability; when the GP signaled high confidence, the AUC for several critical conditions approached 1.0, demonstrating its potential as a clinical decision support tool.
Discussion and Conclusion
DME-GP addresses the "follower" problem common in standard RNNs, where models simply predict the next value will be similar to the last. By using the GP to correct global predictions based on patient history, the model provides more accurate and uncertainty-aware outcomes.
While the model scales linearly with the number of patients, it remains cubic with respect to the number of visits per patient. Future work will explore sparse GP approximations for high-frequency monitoring data and the use of Mixture of Experts to further refine global mean functions for distinct patient cohorts. Ultimately, DME-GP demonstrates that structured mixed-effect models can successfully integrate "big data" population trends with "small data" individual personalization.
Summary updated