Predictive models for cochlear implant outcomes: Performance, generalizability, and the impact of cohort size.
Elaheh Shafieibavani, Benjamin Goudey, Isabell Kiral and 12 others
PMID 34903103WHAT IT FOUND
Machine learning models predicting cochlear implant word scores achieved a mean absolute error of about 20 points.
While modestly better than linear models, this precision is insufficient for individual patient prediction, though the models did identify subgroups with a high probability of substantial improvement.
Key findings
01The best machine learning model achieved a median mean absolute error of 20.81 points in predicting word recognition scores, significantly outperforming linear baseline models.
02External validation showed that models trained on data from two clinics maintained similar predictive performance when tested on a third, unseen clinic.
03Individuals in the highest predicted improvement group had a 94% positive predictive value for achieving a clinically meaningful improvement of 10 points or more.
STILL TO COME
How it was doneWhat they foundWhat it means for SLPs
Read the rest of this summary
You get three full summaries a month, free, and we do not ask for a card. Search, the TL;DRs and your library stay unlimited either way.
What it does not show
The outcome measure (WRS) has inherent test-retest variability, and the follow-up period varied between 6 and 24 months. Data was collected retrospectively from three clinics with different testing protocols, languages, and patient selection criteria. The models relied only on pre-operative data, excluding peri-operative factors like electrode placement or insertion depth which influence outcomes. The study excluded patients with prelingual hearing loss and those with better residual hearing (PTA < 60 dB), limiting applicability to these groups. Self-reported clinical history variables, such as age at onset of deafness, introduce variability.
Declared interests
Funding was provided by the National Institute on Deafness and Other Communication Disorders (N.I.H., Extramural).
The easy way to misread this
Do not interpret the statistically significant improvement of machine learning models over linear baselines as a clinically transformative gain. The difference in mean absolute error is less than one point, and the absolute error remains too high to accurately predict individual patient outcomes.