A blink is not a label: what I am learning from EOG segmentation

Abstract

This fall I have been working with Prof. S. P. Srinivas and Prof. Yuxi Hong on computational problems in optometry. One thread is locating blink events in electrooculographic (EOG) signals: where an event starts, where its electrical peak occurs, and where it ends. This note describes the evaluation questions and review workflow, not private recordings or unpublished performance results. The manuscript and model-development work remain in progress.

I arrived thinking about architectures. I now spend rather more time thinking about what a timestamp means. The timestamp has not thanked me.

1. From samples to events

A temporal model maps an input signal to sample-level scores. Schematically,

$$p_t = f_\theta(x_{t-r:t+r}), \qquad p_t\in[0,1].$$

That is not yet an event detector. A decoder must convert scores into intervals, using rules such as thresholds, minimum duration, and separation between neighboring events. Those rules are part of the model specification, not disposable post-processing.

The work has involved convolutional, temporal-convolutional, recurrent, and U-Net-style approaches. I care less about declaring an architecture winner than about making the whole pipeline inspectable: preprocessing, checkpoint, decoder, and the rule used to compare predictions with references.

2. Detecting a blink and timing it are different tasks

A model can find roughly the right number of events while placing their offsets badly. Conversely, excellent timing on matched events can hide the fact that difficult events were missed entirely.

For a predeclared matching tolerance, detection can be summarized as

$$F_1 = \frac{2TP}{2TP+FP+FN}.$$

For the matched pairs $\mathcal M$, a separate onset error is

$$\operatorname{MAE}_{\mathrm{onset}} =\frac{1}{|\mathcal M|}\sum_{(i,j)\in\mathcal M} \left|\hat t^{\mathrm{onset}}_i-t^{\mathrm{onset}}_j\right|.$$

Both belong in the report. So do extra detections, missed events, offset error, and behavior on verified no-blink intervals. I do not want a metric that rewards the detector for avoiding the hard examples.

Question	What needs checking
Did it find the event?	Matching rule, misses, and extra detections
Did it locate the boundaries?	Onset, peak, and offset errors on matched events
Does it behave when nothing happens?	False detections in reviewed negative intervals
Does the comparison generalize?	Participant grouping and an assessment set not used for tuning

3. A video does not automatically settle the label

The local review tool places EOG and video alongside each other while keeping electrical and optical landmarks separate. An electrical peak is not, by definition, the instant of maximum eyelid closure. Video also introduces frame timing and synchronization questions.

The practical lesson is to record what was marked, in which coordinate system, and by which convention. A model suggestion remains a suggestion until reviewed. A practice annotation is not silently promoted into a certified training label.

This matters for new recordings too. Running a frozen teacher can generate candidates efficiently; it does not manufacture independent human references. The reviewer must be able to add missed events and reject artifacts, not merely approve whatever the model already found.

4. What is implemented, and what remains open

I have worked on offline segmentation and comparison workflows, frozen inference settings, a local EOG/video reviewer, and controlled experiments involving verified negative examples. The ongoing work is establishing a defensible default model and assessment procedure, not announcing clinical readiness.

Repeated recordings, two eyes, and overlapping windows must not be treated as independent people. Where participant linkage is unresolved, the resulting limitation stays explicit. A model selected through repeated validation inspection also needs a separate assessment before I can call its result independent.

Parallel ocular work includes pharmacokinetic modeling questions and molecular-simulation preparation. Structure, stereochemistry, and parameter checks are progress, but they do not establish binding efficacy or treatment benefit. That distinction is as important as any choice of neural network.

5. The connection to my other work

The connection to finance is methodological, not a claim that eyes behave like markets. In both cases, preprocessing and evaluation can quietly answer an easier question than the one we meant to ask. My factor-model work and quantum simulation study keep returning to the same discipline: define the target, preserve the comparison, and report the failure cases.

References

1. Ronneberger, O., Fischer, P., and Brox, T. (2015). U-Net: Convolutional Networks for Biomedical Image Segmentation. Architectural background; the original study concerns images, not this EOG task.
2. Bai, S., Kolter, J. Z., and Koltun, V. (2018). An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling.
3. GROMACS user guide. General reference for the separate molecular-simulation preparation work.

Research methods update, November 2026. No private lab recordings, participant information, or unpublished performance tables are included.