Preoperative prediction of postoperative hypocalcemia after endoscopic thyroidectomy

Reproducible analysis for the ICTMHS 2026 paper: Development and internal validation of a preoperative prediction model for postoperative endoscopic thyroidectomy hypocalcemia.

⚠️ Data notice: Thyroid_Sx_research_(ReadyToClean).csv is patient-level clinical data. Do not keep it in a public repository — see Data & privacy.

Study design

Item	Value
Design	Retrospective cohort, single centre
N	78 patients
Outcome	Postoperative hypocalcemia = postoperative serum calcium < 2.1 mmol/L (10 events, 12.8%)
Preoperative predictors	Pre-op PTH, pre-op calcium, pre-op vitamin D
Candidate models	Standard logistic regression · Firth logistic regression · LASSO logistic regression
Internal validation	Bootstrap (B = 1000), Harrell optimism correction

Events-per-variable (3 predictors) ≈ 3.3 — a small-sample / rare-event regime where penalised methods (Firth, LASSO) are specifically indicated over plain maximum-likelihood logistic regression.

Headline result

After bootstrap optimism correction, all three models fell from an apparent AUC ≈ 0.64 to ≈ 0.53 (95% CI crossing 0.5) — none discriminates better than chance once overfitting is removed. Firth logistic regression was selected as the final model (the preferred estimator for small-sample rare-event logistic regression, with the tightest coefficient CIs). Pre-op calcium was the strongest predictor in every model.

This is a transparent, hypothesis-generating limited result — the appropriate posture for a small preliminary cohort, and stronger scientifically than an overfit AUC that would not replicate. Full interpretation, calibration, decision-curve analysis, and line-by-line manuscript-update guidance are written into the notebook.

Repository contents

Path	Description
model_analysis.ipynb	Main deliverable — the full 7-step analysis (code + per-step "learnings & verdict" markdown for the clinical team)
Thyroid_Sx_research_(ReadyToClean).csv	Raw dataset (patient-level — see Data notice)
old_manuscript_TOETVA_2026.pdf	Current manuscript draft (scope changed TOETVA → endoscopic; many [PENDING] fields)
figures/	Publication figures: ROC (fig1_roc), calibration (fig2_calibration), decision-curve analysis (fig3_dca)
pyproject.toml, uv.lock	Environment (Python ≥ 3.13, managed with uv)

Files beginning with _ (_eda.py, _analysis_dev.py, _build_notebook.py, _sota_metrics.py, _test_sota.py, _manuscript_text.txt) are development scaffolding used to build the notebook; they are not needed to read or rerun the analysis.

How to run

Requires uv and Python ≥ 3.13.

uv sync                       # install dependencies
uv run jupyter lab            # open and run model_analysis.ipynb

Re-execute the whole notebook headless (regenerates all outputs + figures):

uv run --with nbconvert --with ipykernel \
  jupyter nbconvert --to notebook --execute --inplace model_analysis.ipynb

All randomness is seeded (SEED = 42), so results are deterministic and reproducible.

Methods & tooling

• Firth logistic regression implemented dependency-free (Firth 1993; Heinze & Schemper 2002) — the firthlogist package is incompatible with numpy ≥ 2.
• Optimism-corrected AUC via a hand-coded Harrell/Steyerberg bootstrap; cross-checked with DeLong AUC 95% CIs (Sun & Xu 2014), an independent asymptotic method. Generic bootstrap libraries don't implement the fit-on-resample / test-on-original optimism step, and there is no mature Python equivalent of R's rms::validate().
• Calibration: binned observed-vs-predicted, Brier score, and the Integrated Calibration Index (Austin 2020).
• Decision-curve analysis via dcurves.
• A gradient-boosting comparator is shown only as a memorisation check (it hits apparent AUC = 1.0), empirically justifying the regression-only design at EPV ≈ 3.3.

Key references: van der Ploeg 2014 (EPV / data-hungry models); Christodoulou 2019 (ML vs logistic for clinical prediction); Harrell, Regression Modeling Strategies; Steyerberg, Clinical Prediction Models (2nd ed.); TRIPOD+AI (2024).

Data & privacy

The CSV contains patient-level clinical variables (demographics, operative details, labs, surgery dates). Even de-identified, this can be re-identifiable. Do not keep it in a public repository. Make the repo private and/or remove the file from git history, and confirm what your ethics/IRB approval permits regarding data sharing before any publication or data deposit.

Status

• ✅ Analysis complete — model_analysis.ipynb executes end-to-end with 0 errors.
• ⏳ Manuscript [PENDING] fields still to be filled (abstract numbers, inclusion/exclusion criteria, institution, dates, ethics statement).
• 🔜 Recommended future work: larger multicentre cohort, inclusion of surgical-extent variables (which vary little in this mostly-unilateral-lobectomy cohort), and external validation.