README.md

metadata

title: PPMI Mini
tags:
  - medical-imaging
  - mri
  - nifti
  - parkinsons
  - ppmi

PPMI Mini

A 1,000-subject PPMI sMRI evaluation subset derived from local BIDS T1w images and PPMI subject-characteristics metadata.

Cohort

• 1,000 subjects / 1,000 scans
• One T1w scan per subject
• Sex: 431 Female / 569 Male
• Diagnosis: CN 226 / PD 357 / Prodromal 356 / SWEDD 61

Features

The dataset has a single eval split with embedded NIfTI image bytes in nifti. Brain masks and SynthSeg volumes are intentionally omitted for v0.1.

Clinical targets

clinical/ppmi_mini_clinical.parquet holds derived clinical targets for the same 1,000 scans, joined from PPMI LONI study data. It is a sidecar table rather than extra columns on the split, so the imaging is untouched.

from datasets import load_from_disk
from huggingface_hub import snapshot_download, hf_hub_download
import pandas as pd

data = load_from_disk(snapshot_download("medarc/ppmi-mini", repo_type="dataset"))["eval"]
clinical = pd.read_parquet(
    hf_hub_download("medarc/ppmi-mini", "clinical/ppmi_mini_clinical.parquet", repo_type="dataset")
).set_index("sample_id")
clinical = clinical.loc[list(data["sample_id"])]   # aligned to the split

Keys

Column	Meaning
sample_id	Joins to the split. One row per scan.
participant_id	Subject, sub-<PATNO>. Use it to group by subject in cross-validation.

The raw PPMI PATNO is not included; participant_id already encodes it.

Longitudinal targets

Each instrument below yields three columns: <prefix>_baseline, the score measured nearest the scan; <prefix>_slope_48m, the annualized OLS slope over visits in a −0.25 to 4.0 year window relative to the scan, requiring at least 2 visits at distinct dates; and <prefix>_n_visits, how many visits fed the fit. A slope is null where the participant has too few visits in the window.

Prefix	Instrument	Range	Direction	n (baseline / slope)
np3tot_off	MDS-UPDRS Part III motor exam, OFF-medication and untreated exams only	0–132	higher = worse	810 / 765
np3tot_all	MDS-UPDRS Part III, all medication states	0–132	higher = worse	992 / 953
nhy_off	Hoehn & Yahr stage, OFF-medication and untreated	0–5 ordinal	higher = worse	809 / 763
nhy_all	Hoehn & Yahr stage, all medication states	0–5 ordinal	higher = worse	991 / 951
np2ptot	MDS-UPDRS Part II, patient-reported motor experiences of daily living	0–52	higher = worse	998 / 971
mseadlg	Modified Schwab & England ADL, percent independence	0–100	higher = better	849 / 812
mcatot	Montreal Cognitive Assessment	0–30	higher = better	992 / 921
cogcomp	Cognitive composite, see below	SD units	higher = better	987 / 930
sbr_striatum	DAT-SPECT striatal binding ratio, whole striatum, referenced to occipital white matter	continuous	lower = more dopaminergic loss	712 / 440
sbr_putamen	Same, putamen, where loss appears earliest in PD	continuous	lower = more loss	712 / 440
sbr_caudate	Same, caudate	continuous	lower = more loss	712 / 440

MDS-UPDRS Part III is scored either ON or OFF dopaminergic medication, and ON scores are drug-suppressed, so the _off and _all variants are both provided: _off is the cleaner target, _all has the larger sample.

cogcomp is the mean z-score across five full-length neuropsychological tests — HVLT-R total recall, Symbol Digit Modalities, Benton Judgement of Line Orientation, Letter–Number Sequencing and semantic (animal) fluency — requiring at least 3 of the 5 present at a visit. Each test is z-scored across all study visits before averaging. Raw totals are used rather than PPMI's derived age-normed T-scores, so that age is not silently removed from the target.

It is provided because MoCA ceilings in this cohort: 66.7% of participants score ≥27/30 and 12.2% sit at exactly 30, so MoCA cannot resolve variation in the normal range.

Cross-sectional targets

Column	Meaning	Range	Direction	n
upsit_baseline	University of Pennsylvania Smell Identification Test, 40 items	0–40	higher = better	386
rbdsq_baseline	REM Sleep Behaviour Disorder Screening Questionnaire	0–13	higher = more symptoms	990
saa_positive	CSF alpha-synuclein seed amplification assay	1 = positive	—	910
prs_meta5	PD polygenic risk score, META5	continuous	higher = more risk	827
prs_meta5_excl_lrrk2_gba	Same score with the LRRK2 and GBA loci removed	continuous	higher = more risk	827

saa_positive has no specimen collection date in the source, only a visit code and an assay run date, so unlike every other column it is not time-aligned to the scan. One status per subject is taken, preferring the baseline visit. Inconclusive results are dropped rather than coerced. Positivity by group is CN 5.6%, Prodromal 51.3%, PD 90.3%, so it is close to a diagnosis label.

PPMI ships the RBDSQ as individual items rather than a total, so rbdsq_baseline sums items 1–12 and adds 1 for item 13 if any neurological condition is present; note that PTCGBOTH in that source file is not a score but a record of whether the participant, the caregiver or both completed the form. And NHY values of 101 in the source are an out-of-range "not assessed" sentinel, mapped to null here.

Recruitment covariates

Column	Meaning
enrolled_genetic_cohort	1 if enrolled via LRRK2, GBA, SNCA, PINK1 or PRKN carrier status (149 of 1000)
enrolled_hyposmia	1 if enrolled via the hyposmia pathway (197)
enrolled_rbd	1 if enrolled via the RBD pathway (77)

These are covariates, not targets. PPMI recruits genetic carriers through a dedicated cohort, and that membership correlates with PD polygenic risk at r = 0.58 by construction, so any target that might vary with recruitment should be checked against these. The two PRS variants are shipped together for the same reason, to separate cohort enrichment from the rest of the polygenic signal.