Demetrios Agourakis

final-year med student & researcher. computational psychiatry x complex systems. measurement, uncertainty, and models of mind.

[Demetrios Agourakis @DAgourakis]

Every programming language lies to scientists.

Python. Julia. Rust. All of them treat uncertainty as somebody else's problem.

9 months ago I started building one that doesn't. Alone. While finishing medical school.

It now compiles itself. What 2.2M lines taught me — a thread: 🧵

[Demetrios Agourakis @DAgourakis]

@TheColeLab Thanks, and for the 2019 paper. Then my rerun is testing the right thing: matched, 2x, 4x and full rest vs task FC. I'll share the notebook either way. If some subjects do close the gap, would you look at motion or rest state first?

[Demetrios Agourakis @DAgourakis]

@alskaf1293 Thanks, so incremental PC is the right baseline. Does the hardware speedup hold when you compare energy per weight update rather than per step? Happy to read the rewrite when it's out.

[Demetrios Agourakis @DAgourakis]

@TheColeLab @DarwinCompPsych Good point: the evoked mean is already out of task FC, so split-run would remove it twice. Is there a residual path you'd still worry about, like trial-level evoked variability that survives mean removal and is shared by the activations and the FC?

[Demetrios Agourakis @DAgourakis]

@TheColeLab That helps, thanks. In my rerun task FC lifted R² from 0.49 to 0.60, but between-visit errors barely shrank. Happy to redo it at matched rest length and share the notebook if useful to your group. Would you expect more rest data to close that gap, or only raise R²?

[Demetrios Agourakis @DAgourakis]

Oral-only PK data can't separate clearance, volume and bioavailability: you get CL/F and V/F, and flip-flop can swap ka and ke. One early oral sample or an IV arm changes that. Worth checking before a dose recommendation leans on F. #pharmacometrics https://t.co/V4MTFZMPX6

[Computation MDPI @ComputationMDPI]

💊 Highlighted Paper in #Computation!

Explore a mathematical model of medical THC pharmacokinetics following oral administration, offering insights into drug behavior and dosage optimization.

📖 https://t.co/XvUIazwvA5
@KMUTT

#Pharmacokinetics #THC

[Demetrios Agourakis @DAgourakis]

@micahgallen Heart–brain DCM has a confound brain-only DCM lacks: cardiac pulsatility enters BOLD directly. How does the model keep heartbeat-driven vascular signal from reading as cardiac→insula connectivity? RETROICOR before inversion, or a separate vascular observation model?

[Demetrios Agourakis @DAgourakis]

@PTenigma Inverting it: δT ≈ (T²/2256.74)·δD/D, so at 310 K a 1% error in CSF diffusivity is ~0.43 °C, and ±1 °C needs D to ~2.3%. Is ventricular ADC that repeatable across sessions with pulsation and partial volume, and was per-voxel temperature uncertainty reported?

[Demetrios Agourakis @DAgourakis]

@aymannadeem Your fold is the classic case for properties over types: `foldl' (flip (:)) [] xs == reverse xs` states the intent the type can't. In scientific code, units in types catch mm vs m, but no type says 'this is the right model'. What belongs in types, and what in properties?

[Demetrios Agourakis @DAgourakis]

@knzw783 Both routes raise interoceptive weight in choices, so from behaviour alone they may trade off. What separates them in the preprint: a manipulation of cardiac-signal reliability (should move precision) vs one changing the cost of ignoring it (should move preferences)?

[Demetrios Agourakis @DAgourakis]

@ComputationMDPI @KMUTT With oral-only data CL, V and F aren't separately identifiable (only CL/F and V/F), and slow erratic absorption invites ka–ke flip-flop. Did the model fix F from literature or estimate apparent parameters? And was ka checked with a profile likelihood?

[Demetrios Agourakis @DAgourakis]

@ylecun @TimDarcet @gabriberton With a deterministic predictor, multimodal futures collapse to a mean in embedding space unless a latent z carries them, which hurts planning once a is an intervention. Calibrated uncertainty over enc(y)? Energy as a split-conformal score gives coverage without a likelihood.

[Demetrios Agourakis @DAgourakis]

That makes the ID a proxy for 'LLM-review passes', which sits on two paths at once: a mediator for real quality and a direct route to reviewer approval if reviewers run similar models. Testable? Early submissions with v1 identical to the final vs ones revised after an LLM pass would split them.

[Demetrios Agourakis @DAgourakis]

@Xiaodon88925961 @micahgallen So LIP as a relay rather than the source of confidence. In SUR, does the returning processed signal predict single-trial confidence beyond the evidence LIP already carries? That would separate relay from readout.

[Demetrios Agourakis @DAgourakis]

@WeigangCui94 Thanks, Weigang, I'll get back to you once I've run it. Quick one before then: which normative connectome are the lesion networks built on? I'd like to match it so any difference comes from the method, not the template.

[Demetrios Agourakis @DAgourakis]

@whoami3339 Thanks. Did the decomposition stay stable across ICA runs or seeds? That would make the low overlap even more convincing.

[Demetrios Agourakis @DAgourakis]

@WeigangCui94 That's a clean way to frame it: the null sets the hub baseline, and the signal is what rises above it. Looking forward to the paper. Would you share the preprint when it's out? I'd like to try it on a psychiatric lesion-symptom set.

[Demetrios Agourakis @DAgourakis]

@cdickey_ @PNASNews Thanks. If the paired montage happens, trial-to-trial latency jitter is what I'd watch: conduction should give tight jitter, local recruitment a broader spread. Happy to help on the analysis side if useful.

[Demetrios Agourakis @DAgourakis]

@WeigangCui94 Clear, thanks. If the hub suppression comes from the null itself, does it also cost sensitivity for symptoms whose true network runs through hubs? A spin-style null on the same lesions would show how much the anatomy constraint is buying.

[Demetrios Agourakis @DAgourakis]

@TheColeLab Subject-specific task FC from the other visit raised R² from 0.49 to 0.60. Propagated/observed error was 0.71 with independent errors, 1.04 with a session factor. The factor uses the same activations across FC sources. Would split-run activations be a cleaner test?