AJILE12: Long-term naturalistic human intracranial neural recordings and pose

Scientific Data 9:184, doi 10.1038/s41597-022-01280-y; open access (CC BY 4.0), PMC9023453. Read for research direction R1, 4 October 2026. Provenance: papers/amadeus/peterson2022_ajile12.provenance.json.

What was read

• Read in full: the Europe PMC full-text XML converted to text: abstract, background, Methods, Tables 1–5, all figure captions, Data Records, Technical Validation, Usage Notes, statements and references.
• Not read: the only supplementary file, a Human Data Submission Checklist. Figures were not inspected as images. Wang et al. 2016 (Frontiers in Human Neuroscience), the earlier unsupervised-decoding paper on similar recordings, was not read.

What the dataset is

• Who: 12 people in clinical epilepsy monitoring at Harborview Medical Center, Seattle (8 men, 4 women, aged 29.4 ± 7.6), with electrocorticography (ECoG) placed for clinical need. They were chosen because they were active and had electrodes near motor cortex. University of Washington IRB; written consent.
• How long: monitoring lasted 7.4 ± 2.2 days per person, with 8.3 ± 2.2 breaks each lasting 1.9 ± 2.4 h. Only days 3–7 after implantation were kept, to avoid post-surgical anomalies, giving 3–5 days per person, 55 days and 1,280 hours in total. Dates were stripped for de-identification.
• Signals:
  • ECoG: recorded at 1 kHz from at least 64 electrodes per person (surface grids and strips, plus depth electrodes in five people). Released at 500 Hz after drift removal, discontinuity zeroing, a 1–200 Hz band-pass, a 60 Hz notch and common-median re-referencing. On average 7.3 ± 5.6 electrodes were bad on day one; 92.3% ± 6.3% were kept.
  • Video: clinical video at 30 frames per second, about 118 million frames.
• Labels, mostly automated:
  • Pose: for each person, a DeepLabCut network was trained on 1,000 hand-annotated frames (about 0.006% of that person's frames) and applied to every frame, giving 9 upper-body keypoints. Hold-out error was 5.71 ± 1.90 pixels, with 3 pixels about 1 cm.
  • Wrist movements: found by a hidden semi-Markov model on the contralateral wrist, filtered, and visually checked. 6,931 validated events, each with 10 features: reach size, duration, angle, speed, time of day, speech presence estimated from audio, and whether the movement was bimanual.
  • Behaviour: coarse labels from manual review of the video about every 3 minutes. Activity labels are sleep/rest, inactive, talk, TV, computer/phone, eat and other. A "blocklist" marks data breaks, camera movement or occlusion, unrelated experiments, private time, tethering, hands under the blanket and clinical procedures.
• What the hospital days contained (Table 3): sleep/rest was the largest activity label in 11 of 12 people, from 11.9 to 87.3 h per person. The exception, P03, talked for 21.8 h against 11.9 h of sleep/rest and had a 24.4 h data break. Talking ran from 1.5 to 28.3 h and TV from 0 to 22.0 h per person. Private time was 1.1–7.0 h per person.
• Access: the DANDI Archive, dandiset 000055 (version 0.220127.0436), in NWB format. One file per person-day, each a full 24 hours from midnight to midnight. A browser dashboard streams segments without downloading. Code is public.

Main results (verified)

This is a dataset descriptor; the validation is technical:

• Stable spectra: power spectra projected to four regions are consistent across 30-minute windows and across people (on average 160 windows, 80 hours, per person).
• Pose accuracy: the pose models are accurate to about 2 cm on held-out frames.
• Earlier use: the authors' earlier papers used the data to decode naturalistic wrist movements and study their variability.

Limits

• Patients, not healthy people. Recordings come from days 3–7 after brain surgery, during medication changes and seizure monitoring, while confined to a hospital bed.
• Electrode coverage is clinical and was selected here for motor cortex; it varies across people.
• The labels describe observable behaviour: posture, wrist movement, talking, TV, eating. Nothing labels thought, mood or memory. No experience sampling.
• Coarse labels have 3-minute resolution and were made by one annotator. No inter-rater reliability is reported.

What it means for Amadeus

• What clinical monitoring offers (verified): about a week of near-continuous intracranial recording per person, of which 3–5 days are usable. The days are dominated by sleep and rest in a hospital bed, with long gaps. The natural behaviour available is talking, TV, phone and eating. That is the realistic upper bound for "continuous intracranial recording of a person's ordinary life" in clinical monitoring. It is not a month, and it is not ordinary life.
• How unlabelled time was labelled (verified): video, with machine pose estimation seeded by a few hundred to a thousand hand-labelled frames; audio for speech presence; and a human pass every 3 minutes for coarse activity. Private time was marked and excluded, which is a design precedent for a participant's right to unrecorded periods. Video labels what a person does. For what they think, the label still has to come from the person (experience sampling), and this dataset has none.
• E3 (verified for content and access; licence not checked here): public, large and well documented, but with no recall, no thought reports and no later outcomes. It does not fit E3's increment test. Its plausible use is R3's: pretraining or testing a model of the natural dynamics of one brain on days of unlabelled intracranial data (inference).
• Scale reference (verified): 1,280 hours from 12 people is "the largest human neurobehavioral dataset that is publicly available" by the authors' account. A month of one person (about 720 hours) would be a single-person dataset of half that size.