Abstract
We present the first framework for Gaussian-process-modulated Poisson processes when the temporal data appear in the form of panel counts. Panel count data frequently arise when experimental subjects are observed only at discrete time points and only the numbers of occurrences of the events between subsequent observation times are available. The exact occurrence timestamps of the events are unknown. The method of conducting the efficient variational inference is presented, based on the assumption of a Gaussian-process-modulated intensity function. We derive a tractable lower bound to alleviate the problems of the intractable evidence lower bound inherent in the variational inference framework. Our algorithm outperforms classical methods on both synthetic and three real panel count sets.
Introduction
In many empirical studies, particularly longitudinal clinical trials, researchers encounter "panel count data." Unlike recurrent event data where exact timestamps are recorded, panel count data only provide the number of events occurring between discrete observation points. For instance, a patient might report three new symptoms between visits at month 2 and month 4, but the precise timing of these symptoms remains unknown.
Standard temporal point processes are designed for exact timestamps. Modeling panel counts requires inferring a latent intensity function $\lambda(t)$ from interval-censored data. While frequentist methods like LocalEM exist, they often lack robust uncertainty quantification. This paper introduces GP4C, a Bayesian framework using Gaussian Processes (GPs) to model these hidden event rates with full uncertainty estimates.
Methodology: The GP4C Framework
The GP4C (Gaussian Process-modulated Poisson Process for Panel Count data) model assumes that events are generated by a Poisson process with an intensity function $\lambda(x)$. To ensure non-negativity, the intensity is modeled as the square of a latent function $f(x)$, where $f \sim \mathcal{GP}(0, k(x, x'))$.
Tractable Variational Inference
The primary challenge in applying GPs to Poisson processes is the intractability of the Evidence Lower Bound (ELBO). The standard ELBO involves an expectation of the log-integral of the intensity, which cannot be computed analytically.
The authors overcome this by deriving a novel, tractable lower bound. By introducing a hyperparameter $b$, they create a "lower bound of the lower bound" that is differentiable and easy to calculate and optimize. Analysis shows that the choice of $b$ significantly affects the variance of the approximation.
To maintain computational efficiency for large datasets, the model employs sparse GPs with pseudo-inputs (inducing points), reducing the complexity from $O(N^3)$ to $O(NM^2)$, where $M$ is the number of inducing points.
Experimental Results
The model was rigorously tested against synthetic datasets and three real-world clinical datasets: Bladder Cancer, Nausea, and Skin Cancer.
Bladder Cancer Study
The Bladder Cancer dataset tracks tumor recurrences in 85 patients over 53 weeks. GP4C demonstrates a significant improvement in test log-likelihood over the traditional LocalEM method. Crucially, GP4C provides credible intervals, allowing clinicians to see not just the predicted recurrence rate, but also the model's confidence in that prediction.
Key Insights
- Uncertainty Quantification: The Bayesian nature of GP4C allows it to express doubt in regions with sparse data or high volatility, a critical feature for medical decision-making.
- Handling Heterogeneity: An extension of the model, GP4CW, incorporates individual subject weights to account for varying baseline risks among patients.
- Robustness: GP4C is particularly effective at modeling subjects with high event counts (outliers), where frequentist methods often under-fit.
Conclusion
GP4C represents a significant advancement in modeling temporal events from interval-censored data. By providing a tractable variational inference framework for GP-modulated Poisson processes, the authors offer a tool that is both more accurate and more informative than previous state-of-the-art methods. Future work includes scaling the algorithm using Stochastic Variational Inference and extending the framework to spatial point processes.
Summary updated