Data for Human-AI Collaboration for Co-Managing Epilepsy in Real Time

Haoer Shi, M.S.
,
Zack Goldblum
,
William K Ojemann, B.S.
,
Carlos Aguila
,
Kerry Nix
,
Katherine Walsh
,
Sarah Lavelle
,
Kathryn A Davis
,
Nishant Sinha
,
Erin Conrad, M.D.
,
Brian Litt

Raw data for the pioneer project and intermediate files for replicating publication results.

Updated on August 13, 2026 (Version 2)

Corresponding Contributor:

Haoer Shi
Dataset Banner Image
597 Files
285.96 GB
No Records

Dataset Overview

Data for Human-AI Collaboration for Co-Managing Epilepsy in Real Time

This dataset supports human-in-the-loop (HITL / CITL) seizure detection model evaluation and patient-specific fine-tuning. Patient folders use PennEPI identifiers (e.g. PennEPI02012).

Top-level layout

Path Status Description
participants.tsv / participants.json Present Participants table and column sidecar at dataset root
raw_data/ Present Source EDFs and event CSVs under edf/ and events/ (per PennEPI ID)
intermediate_data/ Present Derived analysis products: CITL day predictions, fine-tune runs, aggregate metrics
<dataset root>/
├── participants.tsv
├── participants.json
├── raw_data/
│   ├── edf/
│   └── events/
└── intermediate_data/
    ├── citl_results/
    └── results/

participants.tsv / participants.json

Cohort-level clinical and recording summary. Includes fields such as:

  • PID, PennEPIID
  • Sex, Race, Ethnicity, recording_type
  • Stay / enrollment lengths
  • Seizure counts and average duration
  • Annotation and survey-related flags
  • Laterality, localization, seizure-type notes, medications

participants.json documents column descriptions for the TSV.


raw_data/

Source electrophysiology and annotation timelines for all 13 subjects, organized as parallel edf/ and annots/ trees.

raw_data/
├── edf/
│   └── PennEPI*****/
│       ├── PennEPI*****_day01.edf
│       ├── PennEPI*****_day02.edf
│       └── …
└── annots/
    └── PennEPI*****/
        ├── seizure_events.csv
        ├── patient_events.csv
        ├── sleep.csv
        ├── medications.csv
        └── databricks_events.csv
Item Description
edf/PennEPI*****/ Continuous de-identified recordings for that subject
PennEPI*****_dayNN.edf One EDF per hospital / recording day (day01, day02, …)
annots/PennEPI*****/ Annotation CSVs for the same subject
seizure_events.csv Seizure timelines (original columns, e.g. start / stop)
patient_events.csv Patient self-reported seizures or clinical events
sleep.csv Sleep-stage annotations
medications.csv Medication administration timeline
databricks_events.csv Patient annotations for seizures through the human-AI platform

Same 13 PennEPI IDs appear under both edf/ and annots/.


intermediate_data/

Processed outputs from baseline and fine-tuned seizure detection models. Intended for secondary analysis without re-reading the full continuous EEG streams.

citl_results/

Day-by-day clinician-in-the-loop (CITL) evaluation for a subset of patients.

citl_results/
├── baseline_metrics_all_patients.csv
├── finetuned_metrics_all_patients.csv
├── improvement_metrics_all_patients.csv
├── summary_statistics_all_metrics.csv
└── PennEPI*****/
    ├── day1/
    │   ├── predictions_day1_before_day1.parquet
    │   ├── predictions_day1_after_day1.parquet
    │   ├── predictions_day1_after_day2.parquet
    │   └── …
    ├── day2/
    │   └── …
    └── summary/
        └── evaluation_metrics.csv

Cohort-level CSVs summarize baseline vs. fine-tuned performance and improvement across patients.

Item Description
dayN/ Predictions for evaluation day N
predictions_dayN_before_dayK.parquet Model outputs on day N before fine-tuning that uses feedback through day K
predictions_dayN_after_dayK.parquet Model outputs on day N after fine-tuning that uses feedback through day K
summary/evaluation_metrics.csv Per-patient metric summary across days

Parquet prediction tables generally include timestamps / windows and model scores or labels for continuous evaluation along the hospital stay.

results/

Batch fine-tuning and baseline evaluation outputs (non–day-oriented) for all 13 subjects.

results/
└── PennEPI*****/
    ├── base_model/                    # or onset_base_model / related baseline name
    │   ├── metrics/
    │   ├── pred/
    │   └── prob/
    └── PennEPI*****_<base>_<hparams>_<update_rule>/
        ├── metrics/                   # thresholded metric tables (CSV)
        ├── pred/                      # discrete predictions (CSV)
        └── prob/                      # continuous scores / probabilities (CSV)

Files

Root Directory

1 - 0 of 0 files

No files found.

About this dataset

Publishing history

August 7, 2026
Originally Published
August 13, 2026 (Version 2)
Last Updated

Cite this dataset

Shi, H., Goldblum, Z., Ojemann, W. K., Aguila, C., Nix, K., Walsh, K., Lavelle, S., Davis, K. A., Sinha, N., Conrad, E., & Litt, B. (2026). Data for Human-AI Collaboration for Co-Managing Epilepsy in Real Time (Version 2) [Data set]. Pennsieve Discover. https://doi.org/10.26275/XBU9-ULPN

References

Is Source of

Goldblum, Z., Shi, H., Xu, Z., Ojemann, W. K., Aguila, C. A., Long, K., Xie, K., Nix, K. C., Walsh, K., Chang, E., Lavelle, S., Bach, B., Davis, K. A., Sinha, N., Hammer, L. H., Conrad, E. C., & Litt, B. (2026). An Implantable Device that Converses with Patients and Learns to Co-Manage Epilepsy. https://doi.org/10.64898/2026.01.26.26344234