LUCAID — slide-level correlation and clinical-action analyses

Analysis code and anonymized data to reproduce the slide-level result panels of the LUCAID study: agreement of LUCAID's automated readouts with molecular and pathologist reference standards (correlation), and clinical-action concordance (CAC) against an expert-panel–adjudicated reference standard. Each script regenerates the panels of one figure from a small, fully anonymized table.

Figures reproduced

Script	Figure	What
figure5_cellularity_kras.py	5f–h	tumour cellularity vs KRAS VAF (pathologist / LUCAID cell-count / LUCAID nuclear-area), n = 115
figure6_regression_grading.py	6e/f	tissue-compartment regression grading, LUCAID vs pathologist, n = 140
figure8_clinical_validation.py	8a, c, d, e	prospective validation: rater-vs-reference calibration, MAE, per-task and case-level clinical-action concordance
supplementary_figure4.py	S4a, b	inter-rater correlation matrices; MAE (95% CI) + correlation by task
figure6_survival/	6g–n	spatial-TME prognostic analysis — code and input-table schema only (see Figure 6g–n)

Quick start

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

python figure5_cellularity_kras.py
python figure6_regression_grading.py
python figure8_clinical_validation.py
python supplementary_figure4.py

Each script reads its table from data/, writes PNG + PDF + SVG panels (and a stats CSV) to figures/, and prints the key numbers. The four scripts above need only numpy, pandas, scipy, matplotlib and pillow. Tested with Python 3.11.

Layout

lucaid-analysis/
├── figure5_cellularity_kras.py       # Fig 5f-h
├── figure6_regression_grading.py     # Fig 6e/f
├── figure8_clinical_validation.py    # Fig 8a,c,d,e
├── supplementary_figure4.py          # Suppl Fig 4a,b
├── style.py                          # shared figure style (palette, export)
├── common.py                         # thresholds, categories, MAE + bootstrap CI
├── data/                             # anonymized inputs (+ DATA_DICTIONARY.md)
├── figures/                          # generated panels
└── figure6_survival/                 # Fig 6g-n spatial-TME survival analysis

Data and anonymization

All inputs live in data/ and are documented column-by-column in data/DATA_DICTIONARY.md. Every table is fully anonymized: no case, slide, spot or patient identifiers, no dates, no institution names, no free text, and no molecular data beyond the single KRAS allele frequency used in Figure 5. Cases carry sequential surrogate IDs; the five pathologists are coded P1–P5 (consistently across the prospective tables) and the LUCAID readout is labelled LUCAID. The tables are pre-joined, so no identifier-bearing keys are needed to run the analyses.

Statistics

Correlation with the pathologist consensus is reported as Spearman ρ; correlation with the molecular KRAS reference as Pearson r — in both the retrospective Figure 5f–h (n = 115) and the prospective Supplementary Figure 4b KRAS column. Both KRAS analyses exclude cases with 2 × VAF > 100 % (VAF > 50 %, which break the heterozygous-diploid assumption). All correlations are two-sided. Mean absolute error is reported with a 95% bootstrap confidence interval (1,000 resamples; fixed random seed for reproducibility). Clinical-action concordance uses the study thresholds: cellularity < 10 % vs ≥ 10 %; PD-L1 TPS < 1 %, 1–49 %, ≥ 50 %; MET and TROP-2 H-scores < 100, 100–199, ≥ 200 (see common.py).

All prospective analyses (Figure 8 and Supplementary Figure 4) are restricted to the common 70-case prospective cohort — the cases scored for at least one IHC marker — so every task is evaluated on the same patients.

Figure 6g–n (spatial-TME survival)

The Figure 6g–n panels (compartment/cell-type composition, UICC-adjusted hazard forest, Kaplan–Meier curves, and the PD-L1 × TIL quadrant) use per-patient data from an external NSCLC validation cohort — spatial tumour-microenvironment features together with survival endpoints. That patient-level data is not redistributed here, so figure6_survival/ provides the analysis code and the required input-table schema; see figure6_survival/README.md for how to run it against that table.

License

Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) — free for non-commercial academic research use with attribution; commercial use is not permitted. © 2026 Aignostics GmbH. See LICENSE.

Citation

If you use this code or data, please cite the LUCAID paper (citation to be added on publication).