Analysis code for the manuscript Prediction of Individual Speech Outcomes with Cochlear Implants: Limitations and Opportunities for Clinical Translation.
The paper addresses three questions using the Swiss National Cochlear Implant Database (473 adults with bilateral hearing loss):
- Can pre-operative data predict individual post-operative speech scores, and if not, why? Seven algorithms, from linear regression to a tabular foundation model, across four feature-set sizes, five imputation strategies and six validation procedures.
- Can pre-operative data identify patients at risk of a poor outcome? Binary classification of word-recognition scores below 40%.
- Do early post-operative scores carry prognostic information? The six-month score added to the pre-operative set, and clustering of serial scores into trajectory types.
The main finding is that cross-validated R² does not exceed 0.22 in any configuration, and that this reflects the pre-operative data rather than the models: the same features predict the pre-operative score well (R² = 0.48 from audiometry alone). Implantation selectively removes the predictive value of audiometric and speech measures while developmental and communicative variables retain or gain theirs.
analysis/
prediction/ RQ1 — regression across algorithms and feature sets
equalizer/ RQ1 — how the predictor-outcome relationship changes (W1)
risk/ RQ2 — binary classification of poor outcome
trajectories/ RQ3 — six-month prognostic value and trajectory types
validation/ temporal splits, conformal intervals, learning curves
figures/ scripts that produce the manuscript figures
results/ derived tables (CSV) reported in the paper and Supplement
The Swiss National Cochlear Implant Database is managed by the participating
centre. Data access requires ethics approval and is subject to Swiss data
protection regulations, so the patient-level data are not included here.
The scripts expect a file named CI_UNIFIED_DATASET.xlsx in a data/
directory, which is not tracked.
results/ contains the derived tables that underlie the published figures and
supplementary tables; these are aggregate and contain no patient-level
information.
python -m venv venv && source venv/bin/activate
pip install -r requirements.txtTabPFN is fine-tuned on the present dataset; the fine-tuning runs were performed on an NVIDIA DGX workstation and are not reproducible on CPU in reasonable time. All other analyses run on a laptop.
Campi M, Huber A, Goehring T. Prediction of Individual Speech Outcomes with Cochlear Implants: Limitations and Opportunities for Clinical Translation. (under review)
MIT — see LICENSE.