Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
FAPEX: Fractional Amplitude-Phase Expressor for Robust Cross-Subject Seizure Prediction
Authors: Ruizhe Zheng, Lingyan Mao, DINGDING HAN, Tian Luo, Yi Wang, Jing Ding, Yuguo Yu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Evaluated across 12 benchmarks spanning species (human, rat, dog, macaque) and modalities (Scalp-EEG, SEEG, ECo G, LFP), FAPEX consistently outperforms 23 supervised and 10 self-supervised baselines under nested cross-validation, with gains of up to 15% in sensitivity on complex cross-domain scenarios. It further demonstrates superior performance in several external validation cohorts. To our knowledge, these establish FAPEX as the first epilepsy model to show consistent superiority in SASP, offering a promising solution for discovering epileptic biomarker evidence supporting the existence of a distinct and identifiable preictal state for and clinical translation. We conducted empirical investigations to address the following Research Questions: RQ1: How does FAPEX perform in SASP relative to supervised baselines? RQ2: Does self-supervised pretraining improve performance of FAPEX in SASP relative to self-supervised baselines? RQ3: How well does FAPEX generalize to different cohorts (e.g., species, institution)? RQ4: What is the contribution of each design choice within FAPEX? To evaluate the contributions of each component within FAPEX, we conduct comprehensive ablation experiments. |
| Researcher Affiliation | Academia | Ruizhe Zheng1 Research Institute of Intelligent Complex Systems, Fudan University EMAIL Lingyan Mao2 Department of Neurology, Zhongshan Hospital, Fudan University EMAIL Tian Luo4 Children s Hospital of Fudan University EMAIL Yi Wang4 Children s Hospital of Fudan University EMAIL Dingding Han3 School of Information Science and Technology, Fudan University EMAIL Jing Ding2* Zhongshan Hospital, Fudan University EMAIL Yuguo Yu1* State Key Laboratory of Brain Function and Disorders and MOE Frontiers Center for Brain Science, Research Institute of Intelligent Complex Systems and Institutes of Brain Science, Fudan University, Shanghai Artificial Intelligence Laboratory Shanghai 200232, China EMAIL |
| Pseudocode | No | The paper describes methods and architectural components like Fr NFO, APCE, and SCA using mathematical formulations and textual descriptions, along with figures illustrating the architecture, but it does not contain any explicit "Pseudocode" or "Algorithm" blocks with structured steps. |
| Open Source Code | No | Although in-house datasets are only available upon request so far, we present extensive experiments on public datasets. Moreover, the code of our work will be released upon publication. |
| Open Datasets | Yes | We compile 12 benchmarking datasets spanning four species (human, rat, dog, macaque) and multiple acquisition modalities (Scalp-EEG, ECo G, SEEG, LFP) for evaluation, as summarized in Tab. 1. All recordings are resampled and segmented to standardized lengths, then harmonized to 64 effective channels via channel rejection and duplication, enabling consistent input formatting across all models. See detailed descriptions and preprocessing procedures in App. F. Note that we apply channel alignment during preprocessing to facilitate consistent training across diverse datasets for both our model and a broad range of baselines. In short, FAPEX itself is inherently agnostic to the number and configuration of input channels. Table 1: Summary of datasets. The datasets span several species (human, rat, dog, macaque) and acquisition modalities (Scalp-EEG, ECo G, SEEG, LFP). Dataset Confidentiality Species # Subj. Modality # Ch. # Samples Duration SOP SPH ID/IV OOD/EV FMCE Public Human 65 ECo G/SEEG1 64 32,323 4 s 30 s 1 min ! HUP Public Human 73 ECo G/SEEG 64 53,323 4 s 30 s 5 min ! RESPECT Public Human 6 ECo G 64 17,214 4 s 30 s 5 min ! BEIRUT Public Human 6 Scalp-EEG 64 35,941 4 s 1 min 30 min ! ! CTLE-RATLFP Public Rat 7 LFP 64 11,732 2 s 30 s 5 min ! LPIRE Public Rat 15 LFP 64 159,715 2 s 30 s 5 min ! ! CANINE Public Dog 6 ECo G 64 382,278 4 s 5 min 4 hr ! ! |
| Dataset Splits | Yes | Evaluation protocols. All experiments follow a consistent subject-agnostic nested cross-validation (SANCV) scheme. For each dataset, subjects are split into non-overlapping train, validation, and test folds. These folds are randomly permuted to yield multiple experimental runs for RQ1-3. For RQ1, we evaluate in-domain performance with full supervision. For RQ2, we evaluate in-domain performance by supervised finetuning. for RQ3, we evaluate out-of-domain performance on several regimes for our approach and two best-performing self-supervised baselines. We report median and interquartile range (IQR) across runs for: Balanced Accuracy (BA), Sensitivity (SEN), F1, AUROC, AUPRC. We also report Stratified Brier Score to indicate both discriminative and calibration quality.We calculate F1 as the monitoring score as it captures the trade-off between reducing false alarms and maintaining high sensitivity. We adopt the Friedman test as a nonparametric omnibus for statistical significance with Bayesian post hoc comparison. Refer to App. H for details. |
| Hardware Specification | No | The computations were performed on the CFFF platform of Fudan University. NeurIPS Paper Checklist: Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: We provide detailed information in the appendix. |
| Software Dependencies | No | The main paper text does not explicitly list specific software dependencies with version numbers. The NeurIPS Paper Checklist mentions that full implementation details are provided in Appendix H, but this information is not directly available in the provided text. |
| Experiment Setup | No | We evaluate FAPEX across diverse settings spanning supervised learning (RQ1), self-supervised pretraining-finetuning (RQ2), and cross-cohort transfer (RQ3). This section outlines the baseline, evaluation protocols, and other basic implementation setups common to all experiments. See details of training protocols in App. G. Full implementation details are provided in App. H. All experiments follow a consistent subject-agnostic nested cross-validation (SANCV) scheme. For each dataset, subjects are split into non-overlapping train, validation, and test folds. These folds are randomly permuted to yield multiple experimental runs for RQ1-3. For RQ1, we evaluate in-domain performance with full supervision. For RQ2, we evaluate in-domain performance by supervised finetuning. for RQ3, we evaluate out-of-domain performance on several regimes for our approach and two best-performing self-supervised baselines. We report median and interquartile range (IQR) across runs for: Balanced Accuracy (BA), Sensitivity (SEN), F1, AUROC, AUPRC. We also report Stratified Brier Score to indicate both discriminative and calibration quality.We calculate F1 as the monitoring score as it captures the trade-off between reducing false alarms and maintaining high sensitivity. We adopt the Friedman test as a nonparametric omnibus for statistical significance with Bayesian post hoc comparison. Refer to App. H for details. |