# Analysis plan: do higher depressive symptoms slow cardiovascular recovery from an acute stressor?

## Prediction

An account of depression as a deepened attractor predicts slower return to equilibrium after perturbation ("critical slowing down"). Applied to an acute laboratory stressor: in people with more depressive symptoms, heart rate and systolic blood pressure should stay further from baseline late in recovery (given the size of the stress response), and beat-to-beat fluctuations should become more autocorrelated in recovery than at baseline. A secondary question, fixed now, asks whether a fatigue/low-vigour ("load") dimension predicts slow recovery better than a depressed-mood dimension.

Laboratory studies of depression often report blunted rather than prolonged stress responses, and recovery findings are mixed, so the test is risky; reactivity is reported as a secondary outcome.

## Data

OpenNeuro ds003823 (Yoo et al. 2023, Scientific Data; CC0): the pre-intervention session of a heart-rate-variability biofeedback trial (intervention arms irrelevant here). Stress task with continuous ECG and blood pressure at 2 kHz in phases: baseline (about 4 min), PASAT and Stroop stressors (about 2-3 min each), recovery (about 4 min). PASAT physiology is absent for 62 participants (65 of the 72 older adults lack it; 5 participants lack all phases), so the stress window is the Stroop phase, which every participant with data has and which immediately precedes recovery. Depressive symptoms: CES-D, mean of weeks 1 and 2 (both before the intervention); trait anxiety (STAI-T), POMS depression, fatigue and vigour, likewise. Age group (younger, older) and gender from the phenotype files. Files fetched from the OpenNeuro S3 bucket by digest.

## What was seen before registering

participants.tsv and phenotype column names; which phase files exist for each participant; the CES-D distribution (190 non-missing, 68 at or above 16); phase durations, sampling rate and signal ranges for one participant (sub-5001); and good-beat shares and window-usability flags from code/features.py for all participants (190 with files, 2 unreadable, 162 with usable heart rate and 170 with usable systolic pressure in every window). No heart rate, blood pressure or autocorrelation value was related to any questionnaire.

## Features (`code/features.py`)

ECG R peaks and systolic peaks with artefact rules (see code). Windows: baseline from 30 s; stress = the Stroop phase; early recovery 0-60 s; late recovery 120 s to end; recovery autocorrelation window 30 s to end. Per signal (heart rate, systolic pressure): window means; lag-1 autocorrelation of the linearly detrended good-beat series at baseline and in recovery. A signal is analysed for a participant only if every window has at least 80% good beats and enough beats.

## Tests (`code/analyze.py`)

Rank-based regressions with HC3 errors, covariates older age group and female gender.

**Primary (Holm across four, one-sided, predicted positive):**
1. Late-recovery heart rate on CES-D, adjusting for baseline and stress heart rate.
2. Late-recovery systolic pressure on CES-D, adjusting for baseline and stress systolic pressure.
3. Change in heart-rate lag-1 autocorrelation (recovery minus baseline) on CES-D.
4. Change in systolic lag-1 autocorrelation on CES-D.

**Secondary:** each primary adjusted for trait anxiety; early recovery; reactivity (stress minus baseline) on CES-D, one-sided for blunting; CES-D 16 or more versus less (AUC); POMS fatigue, vigour and depression as separate predictors; the load-versus-mood contrast: the standardised coefficient of a fatigue-minus-vigour composite minus that of POMS depression for late recovery, with a bootstrap CI.

## Interpretation

A primary prediction is supported at Holm-adjusted p < 0.05. The deep-well account is supported overall if at least one recovery-level test (1 or 2) and at least one autocorrelation test (3 or 4) are supported.

## Limits stated in advance

Community volunteers, younger and older, not a clinical sample; CES-D is self-report; recovery lasts about 4 minutes, so the test concerns recovery over seconds to minutes; medication is not recorded; beat-level autocorrelation of heart rate is shaped by breathing, which was not controlled; finger blood pressure is relative.
