Using machine learning to identify speech markers of subjective cognitive impairment in breast cancer survivors

Cancer-related cognitive impairment (CRCI) refers to cognitive changes described by cancer patients. While subjective cognitive difficulties in survivors are well screened by questionnaires, CRCI remains seldom diagnosed by neuropsychological tests. New objective approaches are needed to detect CRCI. The goal of this study is to find the speech markers that can best detect CRCI.

Forty-four women who completed breast cancer treatment and 13 controls were asked to narrate a picture-based story. Speech productions were recorded and analyzed using semi-automatic methods. Fourteen speech features were extracted and incorporated in machine learning models to classify CRCI assessed using the FACT-Cog questionnaire. In addition, all participants underwent psychological assessment to control for confounding factors.

The combination of speech-to-silence ratio, mean duration of silent pauses, and mean duration of filled pauses best predicted cancer survivors with CRCI (ROC = 0.74, accuracy = 73.7%). Bayesian Generalized Linear Models further showed that silence in participants' speech depended on whether they had CRCI and reported depression assessed using the HADS. Credibility Intervals of regression coefficients showed only a trend for reported depression (95% CI [- 0.36, - 0.00]), while scores were below clinical thresholds for depressive disorders for all groups.

Long silent pauses in discourse may be a sign of cancer-related cognitive impairment in survivors of breast cancer.

This study demonstrates the benefits of using speech analysis as a fast, ecological, and non-invasive technique for assessing CRCI.