Skip to content

Quality control and reporting example

Quality control should leave visible evidence. This example combines signal activity, timing checks, missingness, design coverage, and generated figures so downstream exclusions and reporting decisions remain auditable.

Core QC objects

import gpbiometricspy as gp

dat = (
    gp.load_kiosk_demo(participants=["synthetic_kiosk_p001"])
    .copy()
    .iloc[:1800]
    .reset_index(drop=True)
)

gsr_quality = gp.audit_gazepoint_gsr_quality(dat, value_column="GSR_US")

activity = gp.audit_gazepoint_signal_activity(
    dat,
    signal_cols=["GSR_US", "HR", "IBI", "LPMM"],
    group_cols=["participant_id"],
)

resets = gp.audit_gazepoint_time_resets(
    dat,
    time_col="TIME",
    group_cols=["participant_id"],
)

Visual evidence

Missingness overview for EDA heart-rate IBI and pupil channels
MissingnessInspect channel availability across the recorded interval.
 EDA signal quality diagnostic
Signal qualityKeep modality-specific quality evidence separate from model residuals.

The expanded gallery also includes a general biometric-quality panel and design/inference diagnostics generated from package functions.

Audit the experimental design

When the dataset contains participant, trial, and condition structure, inspect coverage before modelling:

design = gp.audit_gazepoint_experiment_design(
    dat.rename(
        columns={
            "participant_id": "participant",
            "MEDIA_ID": "trial",
            "interface_complexity": "condition",
        }
    ),
    participant_col="participant",
    trial_col="trial",
    condition_col="condition",
)

coverage_plot = gp.plot_gazepoint_design_coverage(design)

This catches imbalance and sparse cells before they are hidden inside a downstream model.

Build a report-ready evidence set

A useful QC bundle usually contains:

Evidence Why retain it
schema/input identity shows what was actually analyzed
missingness/activity documents available data and flat/dropout patterns
time resets/gaps constrains event and alignment claims
modality-specific QC records signal-level acceptance/review evidence
event/design coverage shows whether intended conditions/trials were observed
QC figures makes review decisions visually auditable
thresholds/windows makes preprocessing reproducible
exclusions + reasons prevents silent sample-size changes
package/backend versions binds analysis to software identity

Do not turn QC into a hidden score

A single composite “quality” number can obscure why a recording failed. Prefer explicit dimensions—missingness, timing, peak plausibility, interval source, event coverage, interpolation burden—unless a validated composite score is scientifically justified.

Scientific boundary

QC describes properties of the recorded data and the declared analysis pipeline. Passing QC does not prove construct validity, psychological interpretation, causal identification, or equivalence between different sensors.

Continue