IdeaWell

Project ideas from Hacker News discussions.

40 percent of fMRI signals do not correspond to actual brain activity

Original Article

Hacker News Discussion

📝 Discussion Summary (Click to expand)

1. fMRI is prone to false positives if statistical corrections are ignored

“The point of the salmon paper is to demonstrate to people ‘if you do your stats wrong, you’re going to think noise is real’ and not ‘fmri is bs’.” – parpfish
“When we published the salmon paper, approximately 25‑35 % of published fMRI results used uncorrected statistics.” – prefrontal

2. Test‑retest reliability and sample‑size problems make many fMRI findings questionable

“Test‑retest reliability of task‑based fMRI is often as low as 0.16–0.88, with an average of 0.50.” – D‑Machine (citing Bennett & Miller)
“Most fMRI studies are under‑powered; you need hundreds or even thousands of participants to get reliable effects.” – D‑Machine

3. The BOLD signal is an indirect, sometimes misleading proxy for neuronal activity

“The BOLD response is correlated to dephasing induced by the oxy/deoxy hemoglobin ratio… it isn’t even necessarily localized to the voxel.” – physPop
“In 40 % of cases the increased fMRI signal corresponds to a decrease in neuronal activity.” – tsimionescu

4. fMRI is frequently mis‑used or over‑hyped in popular media, clinics, and pseudoscience

“Dr. Amen’s clinics charge thousands for SPECT scans… they’re basically reading tea leaves.” – Aurornis
“The headline ‘40 % of MRI signals’ is misleading; it’s only about fMRI, not all MRI.” – kspacewalk2

These four threads capture the bulk of the discussion: statistical pitfalls, reproducibility concerns, biological validity of the BOLD signal, and the danger of over‑interpretation in the public sphere.

🚀 Project Ideas

NeuroStatGuard

Summary

  • A cloud‑based fMRI analysis platform that enforces best‑practice statistical pipelines, automatically applying multiple‑comparison corrections (FDR, Bonferroni, permutation testing) and providing transparent QC reports.
  • Gives researchers a reproducible, audit‑trailable workflow that reduces false positives and aligns with the concerns raised by “dead salmon” and Voodoo Correlations critiques.

Details

Key Value
Target Audience Neuroimaging researchers, clinical trialists, and data analysts
Core Feature End‑to‑end preprocessing, GLM fitting, automated correction, QC dashboards, and reproducible report generation
Tech Stack Python (NiBabel, Nilearn), Docker, Kubernetes, PostgreSQL, React, D3.js
Difficulty Medium
Monetization Revenue‑ready: $99/month per project

Notes

  • HN commenters like prefrontal and SubiculumCode lament the lack of proper corrections; this tool directly addresses that pain point.
  • The audit trail and reproducible reports make it ideal for journal submissions and peer review, sparking discussion on open science practices.

LongiBrain

Summary

  • A longitudinal fMRI analytics service that uses deep learning to model individual brain activity trajectories, improving test‑retest reliability and enabling personalized monitoring.
  • Tackles the frustration expressed by mattkrause and D‑Machine about low ICCs and the need for within‑subject stability.

Details

Key Value
Target Audience Clinical researchers, longitudinal study designers, neuropsychologists
Core Feature ML‑based reliability estimation, adaptive scan scheduling, personalized baseline modeling, and confidence‑interval dashboards
Tech Stack PyTorch, TensorFlow, FastAPI, PostgreSQL, Grafana
Difficulty High
Monetization Revenue‑ready: $199/month per cohort

Notes

  • Addresses the “40 % of increased fMRI signal corresponds to decreased neuronal activity” concern by quantifying individual variability.
  • Provides a practical tool for designing adequately powered longitudinal studies, a topic repeatedly highlighted by sigmoid10 and D‑Machine.

MultiModalFusion

Summary

  • A web platform that integrates fMRI, EEG, and PET data to validate BOLD signals, offering joint visualizations and cross‑modal correlation analysis.
  • Responds to the call for multimodal validation from kspacewalk2, physPop, and D‑Machine.

Details

Key Value
Target Audience Multimodal neuroimaging labs, translational researchers
Core Feature Data ingestion pipelines, joint ICA, time‑frequency alignment, interactive 3‑D overlays
Tech Stack MATLAB, MNE‑Python, BIDS‑Apps, Three.js, Node.js
Difficulty High
Monetization Revenue‑ready: $149/month per user

Notes

  • By allowing researchers to see how BOLD correlates with EEG and PET, it directly counters the skepticism about BOLD’s validity expressed by physPop and D‑Machine.
  • Encourages cross‑disciplinary collaboration, a recurring theme in the discussion.

BrainVizEdu

Summary

  • An interactive educational web app that visualizes fMRI data, demonstrates the impact of statistical corrections, and explains BOLD limitations to clinicians and the public.
  • Meets the unmet need for transparent, accessible explanations highlighted by Aurornis, freehorse, and Herting.

Details

Key Value
Target Audience Clinicians, medical students, science journalists, lay audiences
Core Feature Simulated fMRI scans, step‑by‑step correction tutorials, “what‑if” scenarios, FAQ chatbot
Tech Stack Unity WebGL, Python backend, Flask, SQLite
Difficulty Medium
Monetization Hobby (open source)

Notes

  • Users like Aurornis would appreciate a tool that demystifies fMRI for non‑experts, reducing misinformation.
  • The interactive “dead salmon” simulation could spark discussion on statistical literacy and media reporting.

Read Later