Construction of the automatic Carcinologic Speech Severity Index (C2SI) score

Introduction : The decrease in mortality and the lengthening of the life span following cancer make the sequelae management of the pathology and treatments a priority, The quality of life of patients treated for oral cavity or oropharynx cancer can be impaired because this pathology modifies the communication abilities of the patients due to its location. The assessment of speech disorders is currently based on perceptual assessments, subject to significant variability. The development of automatic speech treatments can optimize this approach. Objective : assess the validity of the different measurement scores of speech disorders, resulting from an automatic signal analysis, in patients treated for upper aerodigestive tract cancer, to build a global automatic score. Material and methods : our study is based on data from the C2SI projet (Carcinologic Speech Severity Index). 87 patients treated for oral cavity or oropharynx cancer, and 42 controlds performed various speech production tasks, targeting vocal production, prosody, comprehensibility, acoustico-phonetic decoding, and intelligibility. The audio recordings of these productions were then the subject of a human perceptive evaluation, but also of an automatic treatment with the aim of determining different scores. Self-questionnaires of quality of life and perception of speech disability were proposed to the participants to study the links between speech disorder ans perceived impact. Metadata about individual, clinical and treatment information were also collected as part of the search for explanatory factors for speech disorder in patients. Results : The severity of the perceptually assessed speech disorder during an image description task depends primarily on performing a surgical treatment. Among all the parameters that can be extracted from an automatic processing of the speech signal, 6 were selected because they are consistent with the data of the literature, they respect the construct validity by discriminating extreme groups (patients and controls : p-value of the Mann-Whitney U test: p 0, 25). A factor analysis confirms their structure in two domains: 2 parameters are part of the "voice" domain (interquartile difference of the fundamental frequency, and amplitude instability), and 4 are part of the "speech" domain (likelihood scores in acoustic-phonetic reading and decoding, row accumulation and anomalous acoustic-phonetic decoding rates). They are more reliable than perceptual evaluations with an intraclass correlation coefficient of 0.69 [0.62; 0.77] for the inter-judge reliability, and a good internal consistency (Cronbach's alphas greater than or equal to 0.90 in the "speech" domain). This led to the construction of an automatic score by modeling these parameters. It has good metric qualities. Conclusion : Automatic speech processing allows to define valid, reliable and reproducible parameters. It remains to test this score automatically on a new patient sample in the external validation framework. A simplification by reduction of tasks may be considered in routine clinical use.