Résumé
Context:Cancer-Related Cognitive Impairment (CRCI) refers to a cognitive decline induced by cancer and its treatments (1). CRCI is characterizedby cognitive difficulties that cover various aspects of cognition (2). Despite the intensity of the complaints, scores at neuropsychologicaltests struggle to fully match the difficulties reported by patients, as they often are within the norm (3). New ecological tools with objectivemeasurement are thus needed to assess CRCI. Speech analysis is a technique sensitive enough to track subtle cognitive decline that caninform on the speaker's cognitive status (4). Unfortunately, the high number of speech variables that can be found in the literaturecomplicates feature selection. The main goal of this study is to find the optimal subset of speech markers that can best detect CRCI.Method:We recruited 44 participants who completed curative treatments for a breast cancer less than a year ago, and 13 matched healthycontrols. A score of less than or equal to 55/72 at the Perceived Cognitive Impairment subscale of the FACT-Cog questionnaire (5) signalsa the presence of reported CRCI (6). We used this cut-off to split participants into three groups: cancer participants with a cognitivecomplaint, cancer participants without complaint, and controls. Participants also underwent a cognitive and psychological assessment.They produced a narrative discourse task from a five-picture sequence with no time limit. We extracted fourteen speech features andperformed machine learning (ML) analyses to identify the features that can best classify participants of the three groups. Eventually, weran a Bayesian Generalized Linear Model to test the influence of fatigue, depression and anxiety on the speech variables.Results:Cognitive screening, anxiety and depression scores were below clinical alert thresholds for all groups. The ML model is 73.7% accurateusing 3 of 14 speech features (speech-to-silence ratio, the mean duration of silent and filled pauses). Precisely, the model best predictscancer participants with a cognitive complaint and controls. By contrast, the model struggles to classify cancer participants withoutcomplaint. The Bayesian model comparison showed that silence is better explained by a model including group and depression score. Itsuggests that silence in participants' speech did not depend on fatigue and anxiety, but on whether they have CRCI and reporteddepression.Discussion:The ML model used two silence features to classify cancer patients with a cognitive complaint from controls. Hence, the presence of longsilent pauses in discourse may be a sign of cognitive difficulties in cancer participants. This study shows the interest of speech analysisfor assessing CRCI. First, the data are easy and fast to obtain with minimum equipment, since story-telling takes no more than a fewminutes. Second, recording speech is not invasive. Cancer patients undergo long evaluation that often multiply heavy medical imaging.Speech production avoids the reminiscence of these stressful time that can induce changes in cognition (7). Finally, language in use is anecological tool that can complement the FACT-Cog questionnaire. Indeed, speech analysis provides implicit information in the way thatpatients are not aware of what they are being assessed for. A limit is that we focused on patients with breast cancer, and features mayvary depending on cancer type. Future research should develop a full-automatized diagnostic tool combining automatic speechrecognition and a trained classifier to ensure in a clinical context.References:1. Hurria A, Somlo G, Ahles T. Renaming "Chemobrain." Cancer Investigation. 2007 Jan;25(6):373-7.2. Yao C, Bernstein LJ, Rich JB. Executive functioning impairment in women treated with chemotherapy for breast cancer: a systematicreview. Breast Cancer Res Treat. 2017 Nov;166(1):15-28.3. Costa DSJ, Fardell JE. Why Are Objective and Perceived Cognitive Function Weakly Correlated in Patients With Cancer? Journal ofClinical Oncology. 2019;37(14):1154-8.4. Ivanova O, Martínez‐Nicolás I, Meilán JJG. Speech changes in old age: Methodological considerations for speech‐based discriminationof healthy ageing and Alzheimer's disease. Intl J Lang & Comm Disor. 2023 May 4;59(1):13-37.5. Costa DSJ, Loh V, Birney DP, Dhillon HM, Fardell J, Gessler D, et al. The Structure of the FACT-Cog v3 in Cancer Patients, Students, andOlder Adults. Journal of Pain and Symptom Management. 2018;55(4):1173-8.6. Van Dyk K, Crespi CM, Petersen L, Ganz PA. Identifying Cancer-Related Cognitive Impairment Using the FACT-Cog Perceived CognitiveImpairment. JNCI Cancer Spectrum. 2020 Feb 1;4(1):pkz099.7. Hermelink K, Voigt V, Kaste J, Neufeld F, Wuerstlein R, Buhner M, et al. Elucidating Pretreatment Cognitive Impairment in Breast CancerPatients: The Impact of Cancer-related Post-traumatic Stress. JNCI Journal of the National Cancer Institute. 2015 Apr 16;107(7):djv099-djv099.