Find out which parts of the recording deserve to be analysed.

A CSV can be complete and still be unusable. Saturated channels, swapped units, Bluetooth dropouts, clock drift, contact changes and protocol errors all arrive as valid numbers. We inspect those failures before selecting filters or features.

The work covers research-use physiological signals, wearable data, acoustic recordings and other continuous sensor streams. Typical inputs include ECG, PPG, EDA, SPL/SPR, IMU, temperature, audio, event logs and device metadata.

• A pilot recorded successfully, but usable duration varies sharply between participants.
• Motion and contact changes occupy the same frequency range as the event of interest.
• Two devices started together and no longer align at the end of the session.
• A high-pass filter removes drift and part of the biological response with it.
• Event counts change when thresholds move by a small amount.
• Device time, laptop time and protocol time disagree.
• Mean signal quality looks acceptable while a subgroup or condition carries most failures.

Acquisition

Sampling rate, quantisation, gain, clipping, anti-aliasing, firmware transformations, channel mapping, reference configuration and unit conversion.

Time

Clock origin, timestamp monotonicity, dropped samples, duplicate rows, resampling, event latency and cumulative drift between streams.

Signal quality

Flat lines, railing, contact loss, abrupt steps, slow drift, mains contamination, motion, packet loss, focus or exposure failure in signal-derived imaging, and sensor-specific quality indices.

Protocol

Expected phases, cue order, event duration, missing markers, operator notes and whether the recorded state matches the analysis label.

Biological or physical plausibility

Ranges, response latency, morphology, cross-channel coherence and whether a correction has manufactured the pattern it was meant to recover.

• Per-channel and per-session QC tables
• Segment-level quality masks with reason codes
• Drift, missingness and saturation analysis
• Corrected event alignment and a documented time base
• Preprocessing code with parameter configuration
• Feature extraction or event detection where justified
• Retained-data summaries by participant, condition, device or batch
• Methods note and handover

Smoothing can hide exactly the failure that needs to be measured. The audit figure should show four layers:

1. raw signal;
2. detected quality failures;
3. protocol events on the corrected time base;
4. retained data or features used downstream.

A ten-second contact loss remains visible as a ten-second contact loss. It does not become an interpolated physiological event.

Phase 1: representative sample

We inspect a small set chosen to include a good recording, a known failure and an ordinary middle case. The client supplies the acquisition notes and expected event structure.

Phase 2: quality model

We define failure categories, thresholds or learned detectors, reason codes and the unit of exclusion. A whole participant should not disappear because one channel failed for thirty seconds.

Phase 3: reproducible processing

The agreed logic is implemented as code that can be rerun on the complete dataset. Summary tables record what changed and why.

Phase 4: review

The output is reviewed against raw examples. Unresolved cases remain marked rather than being silently forced into “clean” or “bad”.

Physiological

ECG, PPG, electrodermal activity, passive skin potential, respiration, temperature and motion context.

Wearable and mobile

Multi-device studies, free-living recordings, compliance, missingness, battery or transmission gaps, session boundaries and event synchronisation.

Acoustic

Waveforms, spectrograms, voice-activity intervals, clipping, room noise, channel mismatch and recording-level QC.

Other sensors

Continuous instrument or IoT streams where the client can provide the acquisition context and physical interpretation. Asset-specific engineering judgement may require a named domain collaborator.

Can you tell whether my pilot is large enough?

We can test data coverage, event counts, between-participant variation and the stability of a proposed estimate. A universal sample-size answer would be fiction; the target, deployment unit and expected effect determine the calculation.

Do you work with data from a specific wearable brand?

Usually, if raw or sufficiently detailed exported data are available. We need the sampling behaviour, timestamp format, device settings and known transformations. A proprietary daily “stress score” offers little material for signal analysis.

Can motion artefacts be removed?

Some can be detected, modelled or corrected. Others are indistinguishable from the signal of interest without another channel, a better reference or a controlled protocol. We label that ambiguity.

Can data remain in our environment?

Yes. Client-hosted work can be scoped for suitable environments. Access method, logging, permitted exports and deletion are written into the proposal.

Is this a diagnostic service?

No. The standard service produces research and engineering evidence about recordings and analysis methods.

Send the acquisition notes before the whole dataset.

Describe the sensors, sampling rates, protocol, approximate duration, known failure and the decision the analysis must support. We will tell you which representative files are needed for scoping.

Scope a signal review

• Audit a sensor dataset before modelling
• Separate drift, timing error and motion artefact

How Services engagements are scoped