Clinically Collected Variables Fail to Reliably Predict Cognitive Functioning in Untreated Glioma Patients

BACKGROUND
Previous research has identified several clinically collected variables (e.g., age and WHO grade) that are significantly related to impairments in cognitive functions before surgery in patients with a glioma. The degree to which clinically collected variables can collectively explain differences in cognitive functioning between patients, however, remains unclear, as most studies did not evaluate the total amount of variability between patients explained by a set of predictors. Furthermore, only linear or rank-based methods with a small number of variables have been used. In this study, we used various machine learning models to quantify the amount of variability in cognitive functioning we can explain using a comprehensive set of variables as identified from the literature that were collected as part of clinical care. The extent to which cognitive functioning can be reliably predicted reveals the degree to which these variables can explain individual differences between patients.

MATERIAL AND METHODS
We trained 11 machine learning models to predict pre-surgical cognitive function in patients with a glioma (n=354) across eight cognitive tests. Included variables comprised sociodemographics, tumor characteristics, medication, presenting symptoms, ASA score, comorbidities, and anxiety and depression scores. Model performance was evaluated on four objectives 1) predicting impairment on at least one cognitive test, 2) predicting the number of tests on which a patient is impaired, 3) predicting impairment for each test separately, and 4) predicting cognitive function for each test separately. Model performance, hyperparameter selection, and feature importance were analyzed and clinical implications were discussed.

RESULTS
Best-performing models demonstrated above-random performance for most objectives. Performance, however, was unreliable with poor predictions for a substantial number of patients on most objectives. Best-performing models were relatively simple and used most variables for prediction while not relying strongly on any specific variable.

CONCLUSION
Our results show that the clinically collected variables included in this study fail to reliably predict cognitive functioning. This indicates that these variables are unable to explain a considerable amount of variability between patients, despite many of these variables being significantly related to cognitive function in previous studies. Our results suggest that a multi-parametric (i.e., more holistic) view of an individual patient may be necessary to explain differences in cognitive functioning, ideally considering imaging markers, biological markers, sociodemographics, and clinical characteristics. Finally, our results stress the need to collect larger cross-center and multimodal datasets to better explain and predict individual differences in cognitive function.