Predicting residual beta-cell function in Type 1 Diabetes, under federation

Analytical work for the federated prediction of pancreatic beta-cell function from autoantibody and clinical profiles, across four independent ImmPort clinical-study cohorts (SDY524, SDY569, SDY797, SDY1737). Presented at BioITWorld 2026 — talk deck: slides/2026BioITWorld_Federated_Learning_v9.pdf, markdown rendering aligned to the notebooks: SLIDES.md.

The biological premise

Type 1 Diabetes is an autoimmune disease in which the immune system progressively destroys the insulin-producing beta cells of the pancreas. The standard quantitative readout of how many functional beta cells remain is the C-peptide released during a four-hour mixed-meal tolerance test; C-peptide is cleaved off one-to-one with insulin and is not confounded by injected insulin, so the area under its concentration curve (C-peptide AUC) is the field-standard measure of endogenous beta-cell output.

The five autoantibodies in the panel — GAD65, IA-2 (IA2IC), insulin (MIAA), ICA, ZnT8 — are immune signatures of the destruction process. The scientific question: how well does the autoantibody profile predict the beta-cell function that survives the attack?

Read the Limitations before the results. The short answer is that the autoantibody panel alone is not predictive; what predictive power exists comes largely from body-size / age demographics, which are a known confound of C-peptide AUC.

Privacy-preserving federation

Each cohort comes from a different institution and cannot be centralised (IRB, data-use governance). The institutions share only model parameters trained locally; a coordinator aggregates them. No subject-level data ever moves.

This repository simulates that faithfully: every model is fit per study, and only coefficient vectors (or, for random forest, the trained forest) are combined. There is no pooling — not in fitting, not in feature scaling (each institution min–max scales its own data), and not in scoring.

The two-stage architecture

Stage 1 — per-study notebooks (Ridge_<study>, Lasso_<study>, RandomForest_<study>). Each institution fits on its own subjects, LASSO determines the features, and each writes its model artifact for the four selected features (weight_kg, GAD65, received_active_treatment, Sex):

• LASSO / Ridge → vectors/<study>_<method>_sel_vector.csv (coefficients)
• Random Forest → models/<study>_rf_sel.pkl (the trained forest)

(In production the LASSO-determined features become the OMOP-selected feature set, harmonised across institutions.)

Stage 2 — federated notebooks (<method>_<study>_federated_w_SDY_coefficients). Each reads the other institutions' artifacts from disk, applies them to this institution's own data, and produces the presentation graphics: solo vs federated, with R² (bootstrap 95% CI) and MSE. Only artifacts cross institutions; subject-level data never moves.

per-study fit ──▶ vectors/*.csv, models/*.pkl ──▶ federated notebook ──▶ figures/*.pdf
   (Stage 1)         (only parameters cross)            (Stage 2)

Treatment provenance (transitive closure)

received_active_treatment is not taken from a free-text column. It is derived by transitive closure over the ImmPort arm files in data/arms/: subject → arm (*_arm_2_subject.txt) → treatment, where the control arm is identified by name (placebo / control) and the treatment arm is the active-drug arm. This corrected a real error: SDY569's extended data coded arms as 1/2, which the naive =="hOKT3" rule mapped to all-untreated; the closure recovers its 6 treated / 4 control split. SDY1737 has only age-group arms (Adult/Pediatric) and so has no determinable treatment.

Key results

A-priori power — what effect is detectable as institutions federate (the honest power framing; we do not report post-hoc power from observed R²):

LASSO feature selection (federated, Panel B excl. SDY1737) keeps four features — weight_kg, GAD65 (negative: more autoantibody → less C-peptide), received_active_treatment, Sex:

Federation from a single institution's view — the small institution (SDY569, N=10) borrowing the larger study's coefficients. Note the wide confidence interval: this is a transfer result on few subjects, not a within-institution power gain:

Reproducing

conda env create -f environment.yml
conda activate oadr-autoantibody
jupyter lab

Run order: Stage-1 per-study notebooks first (they write vectors/ and models/), then the Stage-2 *_federated_w_SDY_coefficients notebooks (they read those artifacts and write figures/). All notebooks are generated deterministically by the scripts in bin/notebook_builders/ (fixed seed 42).

Notebook map

Notebook(s)	Role
Ridge_/Lasso_/RandomForest_SDY{524,569,797,1737}	Stage 1 — per-institution fit, write artifacts
Federated_Lasso	federated LASSO + ADMM, feature selection
{LASSO,Ridge,RandomForest}_SDY{524,569}_federated_w_SDY_coefficients	Stage 2 — apply partners' artifacts, presentation graphics
Apply_Coefficients_SDY{797,1737}	reduced-feature analysis for the studies that cannot fully join
Predicting_Residual_Beta-cell_function	k-axis sweep across all methods (backup)
c-peptide_baseline_to_tidy	data preparation

Superseded / exploratory notebooks are in archive/ipynb/.

Limitations

These are real and material.

1. weight_kg is a body-size / age confound, not beta-cell biology. It is the dominant feature (coefficient ≈ +0.97) and correlates with C-peptide AUC at r = 0.67–0.93 — as do height and age. C-peptide AUC is the standard AUC-mean concentration and is not body-size normalized (the ImmPort ADCPEP files record weight as a separate column). Age-at-onset is the best-known predictor of residual C-peptide, and weight/height proxy it in pediatric cohorts. So the extended panel's predictive power comes from demographics, not the autoantibody profile — and this is why the model fails to transfer to SDY1737 (where weight r ≈ 0). The SDY1737 weights were cross-checked against the authentic ImmPort PHEXMSTR physical-exam file and match exactly, so its null weight–C-peptide relationship is a genuine population difference, not a data error.

2. No post-hoc power. Computing "achieved power" from an observed R² is circular and was removed. Power is reported only a-priori (detectable effect vs N, above).

3. Small N — wide confidence intervals. SDY569 has 10 subjects and 4 predictors; its federated R² point estimate of ~0.50 carries a 95% bootstrap CI of roughly [−0.2, +0.8]. Read the interval, not the point.

4. "Federation gain" is transfer, not power. With only a small institution's own subjects to score on, the federated number shows the larger study's model fitting the smaller study — not that more N raised power within the small institution.

5. received_active_treatment conflates different drugs (anti-CD3 hOKT3 / teplizumab vs anti-CD2 alefacept) as one binary. Valid for the SDY524+SDY569 pair (both anti-CD3); a simplification if generalised.

6. Researcher degrees of freedom. The L1 penalty (α = 0.008) and the cohort (excluding SDY1737) were chosen after observing they work; some optimism is baked in. No external validation cohort exists.

7. Assay heterogeneity. SDY797's autoantibodies are binary (positive/ negative) rather than continuous titres, and GAD65's sign differs across studies — both limit clean coefficient transfer.

Layout

src/oadr_data.py             shared loader (per-study quirks, panels, treatment closure)
data/arms/                   ImmPort arm files for treatment provenance
ipynb/                       active notebooks
archive/ipynb/               superseded notebooks
bin/notebook_builders/       deterministic notebook generators
vectors/ , models/           Stage-1 artifacts (only parameters)
figures/ , results/          Stage-2 outputs
slides/                      BioITWorld 2026 deck (v9 PDF) + SLIDES.md rendering
environment.yml              conda environment