Step 1: ComBat Batch Effect

• Input: Raw network-level
  • calculation: take average of parcel id
  • covariates: age, sex
  • batch effect: site
• Goal: Remove site-related differences while preserving covariate effects (sex, age)
• Method: neuroHarmonize.harmonizationLearn( )

Step 2: Define Cases vs. Healthy Controls

• Goal: Split the dataset into:
  • Healthy Controls (HCs) — for learning ‘normal’ variation
  • Patients (cases) — for progression modeling via SuStaIn
• Exclusion Criteria: Similar with previous study published at Obesity (10.1002/oby.23901)

Data for this project were obtained from the ABCD 3.0 data release (DOI:10.15154/1524739), and they focused on baseline assessments collected from September 1, 2016, to October 15, 2018. Participants were excluded if they met the following criteria: not fluent in English, history of seizures, birth more than 12 weeks premature, birth weight less than 1200 g, complications at birth, substance use disorder, intellectual disability, traumatic brain injury, brain tumor, stroke, aneurysm, brain hemorrhage, subdural hematoma, cerebral palsy, diabetes, lead poisoning, muscular dystrophy, autism spectrum disorder, and other medical conditions considered exclusionary. Additional exclusion criteria were applied based on anthropometric measurements, neuroimaging, and other covariates as described in the following sections…

• Result:

  • HCs: participants without prenatal exposure to GDM, currently in healthy BMI status, additionally reserve those with no adverse cases in KSADS (n=3557).

  • Cases: participants with prenatal exposure to GDM (Gestational Diabetes Mellitus), regardless of currently BMI status (n=574).

    

    

Step 3: Regress Out Covariate Effects (HC-based Residualization)

• Goal: Remove normal biological variation (sex, age, race/ethnicity, pubertal stage, parental education, household income, handedness, eTIV) using HCs only
• How:
  1. Fit linear models: ROI ~ covariates using only HCs
  2. Apply model to all subjects (HC + cases)
  3. Save residuals: actual - predicted
• Do not include site anymore (already adjusted in ComBat)
• Use only HCs to build the model of normative (non-disease) variability

Step 4: Z-score Transformation

• Goal: Standardize all features into z-scores relative to HCs, for comparability across ROIs

• How:

  1. For each ROI, calculate mean_HC and std_HC from HCs only

  2. Apply:

     $z = \frac{\text{residual} - \mu_{HC}}{\sigma_{HC}}$

• SuStaIn expects standardized input:

  Higher z-score → more abnormal, (usually cortical volume decrease → negative z-score)

• No regions were significantly different between groups after the FDR correction.

Step 5: Feed Z-scores to SuStaIn

Heart racing, hands shaking—it’s time to feed the beast… with data!