Epidexia

Research

Acceptance email: Decision: Accept. We agree with the reviewers that this paper is a good fit for the workshop.
Acceptance email · AISciK 2026 program chairs via OpenReview · 29 Sep 2026
01

Split and Aggregation Sensitivity in the ProteinGym Supervised Benchmark

NeurIPS 2026, AI for Science workshop · Accepted

ProteinGym is the standard supervised benchmark for predicting how mutations change a protein's fitness, and leaderboards on it decide which models get used. This paper shows that the ranking of methods depends on choices that are rarely reported: how the data is split and how per-assay scores are aggregated. Under different but equally defensible protocols the same set of models reorders, so a claimed state of the art can be an artefact of the evaluation rather than of the model. Accepted at the NeurIPS 2026 AI for Science workshop at a 22.5% acceptance rate.

Safwaan Majid, Aarnav Trivedi, Arnav Mahajan, Saketh Nandam, Srujan Chilakapati, Pratyay JakkulaProtein fitness, Benchmarks, Evaluation protocols
Read the paperdoi:10.2139/ssrn.7599298
Acceptance email: Accepted for presentation at the IEEE BigData 2026 High School Symposium, 17 December 2026, Phoenix, Arizona.
Acceptance email · IEEE BigData 2026 High School Symposium program committee · 9 Oct 2026
02

Model Selection in Traffic Forecasting under Random Missingness and Sensor Outages

IEEE BigData 2026, High School Symposium · Accepted

Random dropouts and sensor outages can hide the same amount of traffic data while leaving very different information behind, so a model chosen under one failure mode may not be the right choice under the other. On METR-LA this paper holds training and the selected model family fixed, then compares four compact forecasting pipelines across random masks and temporal-block masks. Error is higher under blocks, but the ranking of models never reverses and selection regret is 0.000 in both regimes, separating a loss of accuracy from a change in which model you should prefer.

Safwaan Majid, Aarnav Trivedi, Pratyay Jakkula, Arnav Mahajan, Saketh Nandam, Srujan ChilakapatiTraffic forecasting, Missing data, Model selection
Read the paperdoi:10.2139/ssrn.7599278
Acceptance email: Accepted as a short paper. 2463 submissions, 480 short papers accepted, a 19.48% acceptance rate.
Acceptance email · IEEE BIBM 2026 program committee · 23 Sep 2026
03

A Leakage-Safe Temporal Benchmark for Target-Disease Prioritization: Explaining Network Uplift through Local Support Overlap

IEEE BIBM 2026 · Accepted short paper

A benchmark for ranking which disease targets to pursue next, built so that no future knowledge leaks into training because the temporal split is enforced rather than assumed. Evaluated on it, the gains that network methods show over simpler baselines trace back to a single measurable cause: how much local support a candidate pair already has in the graph. Accepted as a short paper at IEEE BIBM 2026 at a 19.48% acceptance rate.

Aarnav Trivedi, Arnav Mahajan, Srujan ChilakapatiBenchmarks, Network biology, Target prioritization
Read the paperdoi:10.2139/ssrn.7525601
Acceptance email: Accepted with revision. Overall, this is a good paper. Suitable for presentation at ICOCO 2026 after minor revisions.
Acceptance email · ICOCO 2026 via EDAS · 31 Aug 2026
04

Does Adaptive PM2.5 Correction Transfer Across Devices, Sites, and Time? A Prospectively Frozen Three-Role PurpleAir Evaluation

IEEE ICOCO 2026 · Accepted

Low-cost PurpleAir sensors need correction before their PM2.5 readings can be trusted, and adaptive corrections that use a sensor's own history look excellent on the data they were fitted to. This paper asks whether that advantage survives on a new device, at a new site, and forward in time. In a prospectively frozen evaluation of 21 public sensors across 12 sites, with roles and protocol fixed before any test data was seen, the adaptive model's error was 14% to 20% higher than a static correction in every setting, a negative result reported as carefully as a positive one.

Safwaan Majid, Aarnav Trivedi, Arnav Mahajan, Saketh Nandam, Srujan Chilakapati, Pratyay JakkulaAir quality, Sensor calibration, Evaluation protocols
Read the paperdoi:10.2139/ssrn.7525664
Acceptance email: Accepted for oral presentation at IEEE IATMSI 2026, ABV-IIITM Gwalior, 12 to 14 March 2026.
Acceptance email · IEEE IATMSI 2026 via Microsoft CMT · 18 Dec 2025
05

ADNET: Dual-Stream Alzheimer's Diagnosis Network with Integrated Grad-CAM Explainability

IEEE IATMSI 2026 · Accepted

ADNET is a dual-stream convolutional network that stages Alzheimer's disease from a single MRI slice, from cognitively normal through early, mild, and late impairment to AD. By pairing two parallel feature streams with Grad-CAM heatmaps of the regions behind each decision, it reaches 97.55% test accuracy while showing clinicians exactly where it looked, so a prediction can be checked rather than trusted.

Aarnav TrivediDeep learning, Neuroimaging, Explainability
Read the paperdoi:10.36227/techrxiv.175987989.93553390/v1