Skip to content

Align the chance baselines of TOPS-fMRI Pearson-r and AUC subscores #62

Description

@GTC2333

Problem

The TOPS-fMRI score combines two subscores with different chance baselines:

  • Study4 uses Pearson r clipped to [0, 1], where an uninformative constant prediction scores about 0;
  • Study5 uses raw AUC, where an uninformative constant prediction scores 0.5.

The final score is 0.5 * study4_score + 0.5 * study5_score.

Trigger scenario

A constant, uninformative model receives study4_score = 0 and study5_score = 0.5, producing an overall score of about 0.25.

Impact

  • An uninformative model receives a non-zero headline score.
  • The effective score range is compressed.
  • Study4 and Study5 contribute equally despite using different zero points.
  • A TOPS score of 0.5 is difficult to interpret.

Expected fix

In a new TOPS task version, align both subscores to the same chance baseline. For example, transform AUC with max(0, 2 * AUC - 1).

Keep raw Pearson r and AUC values as diagnostic metrics, document the chance baseline and formula, regenerate the baseline for the new task version, and do not retroactively apply the new formula to old results.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions