MBBS. Clinical research in neuro-oncology, moving toward computational methods.

The work here follows one question: oncology trials treat the brain, so how often do they actually measure what happens to it?

Registration is where a trial declares what it intends to measure. Across 93,371 outcome-bearing ClinicalTrials.gov cancer records and 14 functional domains, this asks whether the outcome matching a drug's best-documented toxicity is among them. The anchor is cisplatin, which causes permanent hearing loss in a large share of the patients who receive it.

Of 2,189 modern therapeutic cisplatin trials, 27 register a hearing outcome. 1.72% after direct standardisation for cancer site, phase and era, against 0.15% in matched comparator strata — rate ratio 11.88 (95% CI 6.34–21.92). Corrected for classification error the corpus-wide rate is 0.12% (0.06–0.17) against an observed 0.21%. Relative enrichment and an absolute rate under 2% are both true at once.

The ground truth is 7,966 paired domain labels from two independent clinician passes, 97.21% raw agreement, all 222 disagreements resolved by documented human adjudication. No model output enters it.

Accuracy uncertainty comes from a cluster bootstrap of the validation design rather than independent Beta approximations per cell — sensitivity and specificity are drawn from the same resample, so their dependence survives, and prevalence stays inside [0,1] by construction instead of hitting the boundary the Rogan–Gladen plug-in produces. Rate ratios use BCa intervals; switching from percentile intervals moved two secondary pairs across the null, which is reported rather than smoothed.

One arm is deliberately unweighted. The enrichment sample drew 40 records from a stratum of 71 for hearing, but which records belonged to which stratum lived only in a sampling key that no longer exists. Assigning those records a probability from the wider frame is wrong by three orders of magnitude, so they carry no weight and are reported as an unweighted audit of the predicted-negative space instead.

Cross-sectional registry analysis of 289 randomized glioblastoma and high-grade glioma trials from ClinicalTrials.gov, ISRCTN, EU-CTR and the WHO ICTRP registries.

Objective neurocognitive outcomes appear in 13.5% of trials. Health-related quality of life appears in 37.2%. Paired discordance is 70 trials measuring HRQoL alone against 4 measuring cognition alone (McNemar P<.001). Academic and cooperative-group sponsors register cognition at roughly three times the industry rate.

Firth penalized logistic regression rather than ordinary maximum likelihood, because the international-registry term separates completely at 0/14 events and ordinary GLM returns an odds ratio of 0.000 with an infinite interval. Bootstrap optimism correction, E-values, and a sponsor-by-domain interaction test that comes back at P=.389 — reported, because it limits the claim.

Every table and figure regenerates from one script; the run asserts cohort size and event count before it estimates anything. Manuscript under revision for the Journal of Neuro-Oncology.

The same question at registry scale, across CNS, breast, lung and head & neck.

1,888 trials hand-labelled one at a time rather than keyword-matched, because the distinctions that matter are not lexical: a cognitive instrument counts, the NANO neurologic exam does not, Karnofsky performance status does not, and a quality-of-life questionnaire's cognitive subscale does not.

A TF-IDF + LASSO baseline in R posts AUROC 0.971 and adds zero lift over a naive keyword rule. A fine-tuned Bio_ClinicalBERT appeared to beat that rule for CNS — the one cancer type where the vocabulary is heterogeneous.

Then I checked the input. The outcome text I had trained on was truncated at a median of 400 characters against ~1,760 in the registry, and for most positive trials it no longer contained the instrument. The test is the 113 trials I had labelled positive that keyword flagging missed — the ones that made the task look hard. On the truncated text the keyword rule recovers 1 of 113. On complete registry text, the same unchanged list recovers 113 of 113, and neither model improves on it: the deployed BERT gets 72% recall on CNS where the rule gets 100%.

The transformer was solving a problem my own text pipeline had created. That correction is written into the repository against the original claim rather than edited out of it.

The classifier deployed and serving — hybrid routing, BERT for CNS and the keyword rule elsewhere, because that is what the evidence supported.

Live · FastAPI, Docker, tests in CI. The model is int8-quantized to 169MB from 433MB to survive a 512MB instance, and batches are capped at 3 after larger ones were out-of-memory killed in production.

Every prediction carries a review flag. Text matching a documented failure pattern is flagged even when the model is confident, because on those categories confidence carries no information.

Tools — R (tidyverse, glmnet, logistf, ggplot2) · Python (PyTorch, transformers, FastAPI, pandas) · Docker · registry APIs

khanmubashirahmad@gmail.com