Self-report
"I'm fine"
Flat across every condition. Language under-reports load.
TL;DR
The Body Says Otherwise is the research strand behind a gamified cognitive-training platform built to measure cognitive flexibility in over 200 children. I designed the trials, analyzed the datasets in R, and built a PyTorch prediction model on behavioral and physiological signals.
The model's predictions failed in a revealing way. A child would score well on the task while their body showed distress — elevated heart rate, suppressed alpha-wave activity — and still report "I'm fine." The systematic gap between self-report and physiology became the finding.
200+
children measured
7
experimental trials
55+
studies synthesized
2
teams unblocked
THE PROBLEM
You can measure a task score. You cannot ask a seven-year-old what their nervous system is doing.
Cognitive flexibility — the ability to switch rules, strategies, and frames — is one of the executive functions that matters most for how children learn. The platform, Mate Marote, measures it through gamified switching tasks, historically scored on two variables: accuracy and response time.
Those two numbers describe the output of thinking. They say nothing about its cost. A child can produce a correct answer while paying a physiological price for it — and the standard measures are blind to that price.
So I widened the aperture. Alongside behavior, I recorded physiological signals: heart rate, and alpha-band EEG activity, which drops when the brain is under load. The plan was simple: if self-report and task score both track effort, a model trained on physiology should predict them. It didn't. And the way it broke was the point.
A child would perform well, hands fidgeting, breathing heavier — and say, "I'm fine."
Children in Argentina echoed something I'd later hear again in Berlin: the body was telling a different story than the words. If people under cognitive load can't verbally report their state, perhaps their bodies can compose it instead.
THE SOLUTION
An end-to-end pipeline: design the trials, isolate the drivers, then model the signal underneath the score.
Experimental design & behavioral analysis
-
01
Seven trials, 200+ datasets
Designed and ran seven experimental trials evaluating the cognitive-training product for children, then analyzed 200+ participant datasets in R.
- ANOVA across conditions to isolate which product features measurably drove performance
- Separated real effects from noise before anything reached the model
-
02
A measure that survives switching — CORSWOPS
Beyond accuracy and RT, I evaluated switch ratio (CORSWOPS) — correct switches over opportunities to switch — as a flexibility index less biased by baseline accuracy and age.
- Switch ratio significantly predicted accuracy gains (p = 0.003, R² ≈ 0.36)
- But cross-validation exposed its fragility (mean CV R² negative) — a caution I carried into the model, not a result I oversold
The prediction model
-
03
Behavioral + physiological signals in PyTorch
Built a prediction model in PyTorch on combined behavioral and physiological signals, alongside secondary analysis of human fMRI and eye-movement data.
- Inputs: task behavior, heart rate, alpha-band suppression, gaze patterns
- Target: the child's reported and scored state
-
04
The informative failure
The model couldn't reconcile the inputs — because they genuinely disagreed. The physiology said loaded; the self-report said fine. Rather than paper over it, I surfaced the divergence as the result: a systematic, reproducible gap between what children reported and what their neural and behavioral data showed.
Synthesis that moved the teams
-
05
55+ studies into shared briefs
Synthesized 55+ peer-reviewed studies on cognitive-flexibility measurement into briefs both sides could act on — breaking a deadlock between the research and product teams over which measures the platform should adopt.
THE SIGNAL
What the child said, and what the body showed — at the same moment.
This is the gap the model kept hitting. Switch between a child's self-report and their live physiological trace during the same task window. The words hold steady. The body doesn't.
participant · anonymized · switching task, block 3
Self-report
"I'm fine."
Physiological trace
heart rate
— bpm
alpha-band power
— %
the words don't move — the body does
Alpha suppression
↓ 8–12 Hz
Drops sharply as switch cost rises — an index of effort the child never names.
Model verdict
divergent
The prediction fails precisely where report and physiology disagree.
THE RESULT
The failed prediction became the finding.
- Isolated, via ANOVA over 200+ datasets, which product features measurably drove performance — giving the product team an evidence-based basis for what to keep.
- Surfaced a systematic gap between children's self-report and their neural/behavioral data, reframing "the model failed" as "self-report is an unreliable channel for cognitive load."
- Delivered shared briefs from 55+ studies that broke a research-vs-product deadlock on which flexibility measures the platform should adopt.
- Carried the core question forward: if the body composes a state words can't report, that state can be measured — and, later, made audible.
WHAT THIS PROJECT DEMONSTRATES
-
Experimental design & statistics
Seven trials, 200+ participant datasets, ANOVA-driven feature isolation, and a critical read on a measure's cross-validated limits — reported honestly rather than oversold.
-
Physiological ML modeling
A PyTorch prediction model over behavioral and physiological signals, with secondary fMRI and eye-movement analysis, treating a model failure as a source of insight.
-
Research synthesis
55+ peer-reviewed studies distilled into briefs that aligned two teams with competing priorities around a shared measurement standard.
-
Interdisciplinary framing
A neuroscience result about the limits of self-report, framed so both a product organization and a research lab could act on it.