Session Item

Irradiating brain or head & neck tumor with curative or prophylactic intent, results in dose to the healthy surrounding brain possibly resulting is clinical relevant side effects like neurocognitive decline. Preservation of neurocognitive function is of paramount importance to maintain the quality of life of the treated patients. Prediction models have the potential to identify patients at high risk of developing radiation-induced neurocognitive burden. This study summarizes and evaluates available prediction models for estimating the risk of neurocognitive decline after cranial irradiation.

MEDLINE was searched on 4 March 2021 for publications containing relevant truncation and MeSH terms related to “radiotherapy”, “brain”, “prediction model”, and “neurocognitive impairments” (e.g., memory dysfunction, learning and attention deficits, problem-solving incompetence, psychological disorders). Two independent reviewers excluded studies according the following criteria: lack of model specifications, no predictor in multivariate analysis, or no adult population. Quality of prediction models was assessed using 14 common methodological considerations in machine learning proposed by Andaur Navarro CL et al. (PMID: 33177137), including data source, data preparation, hyper-parameter tuning, model building strategy, test of interaction terms, and applying shrinkage or penalization methods.

Of 3,351 studies reviewed, 27 studies met the pre-defined eligibility criteria. Included studies were published between 1996 and 2021 with a median sample size of 119 and 55.6% male patients. Nineteen studies developed a prediction model for patients with primary brain or head & neck tumors and eight studies included patients with metastatic brain tumors. Four studies assessed the effect of prophylactic cranial irradiation. Hopkins Verbal Learning Test-Revised (n=7, 26%), Montreal Cognitive Assessment (n=3, 12%), and Mini Mental State Examination (n=2, 7%) were the most frequent neurocognitive outcome assessment tools. All studies used regression (n=14 linear, n=8 logistic, and n=5 cox proportional hazard) as the machine learning method. Further details of the included studies are described in Table 1. The median quality score was 2 out of 14 and only one study assessed the area under the receiver operating characteristic curve (Figure 1).

Clinical prediction models may support the risk estimation of neurocognitive decline after brain radiotherapy, but existing models have limited quality and are at high risk of bias. Therefore current prediction models are of limited clinical value. Future studies should adhere to model building standards and conduct model validation to derive a reliable prediction model that can be used in clinical routine for decision support.