Differentiable latent structure discovery for interpretable forecasting in clinical time series
Differentiable latent structure discovery for interpretable forecasting in clinical time series
Authors: Ivan Lerner, Jean Feydy, Alexandre Kalimouttou, Anita Burgun, Francis Bach Date: 2026-04-30 Paper ID: openalex:2604.27967
Summary
The paper introduces StructGP and its extension LP-StructGP for interpretable forecasting in irregular clinical time series. StructGP utilizes continuous-time multi-task Gaussian processes with differentiable structure learning to identify sparse DAGs of variable dependencies, while LP-StructGP incorporates latent pathways to capture cross-patient progression. Empirical results on MIMIC-IV and the PhysioNet Challenge show that these models provide superior calibration and accuracy compared to unstructured kernels and baseline methods.
Key Contributions
- Introduced StructGP, a continuous-time multi-task Gaussian process that performs differentiable structure learning to recover sparse DAGs.
- Developed LP-StructGP, which uses subject-specific coupling filters and gating to capture shared cross-patient progression trajectories.
- Demonstrated superior forecasting accuracy and uncertainty calibration compared to unstructured kernels and independent-task baselines on MIMIC-IV septic shock data.
Open Questions & Future Work
Key Concepts
- structgp: A continuous-time multi-task Gaussian process that uncovers sparse, ordered DAGs of inter-variable dependencies using process convolutions.
Archivist Review
The review prioritized the core contribution of differentiable structure learning within continuous-time Gaussian processes. StructGP was approved as a significant architectural paradigm. LP-StructGP was rejected as a subordinate model extension. One open question regarding the fundamental limitation of Gaussian assumptions in process convolutions was approved, while the scalability concern was rejected as a standard optimization challenge.
Approved Concepts
- StructGP: The model introduces a novel way to combine differentiable structure learning (DAG recovery) with continuous-time multi-task Gaussian processes.
Approved Open Questions
- Non-Gaussian Likelihoods for GPs: This is a fundamental bottleneck for deploying continuous-time Gaussian process models in real-world clinical or physical systems that are not inherently Gaussian.
Rejected Candidates
- [concept] LP-StructGP (
lp-structgp) - subcomponent_of_broader_mechanism: This is a specific extension of the primary StructGP model and does not warrant a separate entry. - [open_question] Scalability of latent pathway GPs (
scalability-of-latent-pathway-gp-models) - low_impact: This is essentially a parameter scaling and efficiency problem which is common in model development, rather than a fundamental research bottleneck.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2604.27967
- paper_id
- 2604.27967
- paper_source
- openalex
- domain
- time-series
- tags
- time-series-forecastinginterpretabilityuncertainty-quantificationelectronic-health-records
- architectures
-
- datasets
-
- concept_slugs
- structgp
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-03T06:02:36Z