ki element’s overarching mission is to use speech to recognize and classify neurological and psychiatric symptoms.
Speech efficiently provides a lot of information about a person’s cognitive, affective or motoric profile. A wide range of motor and cognitive skills are required to produce speech (for more information about language processing read one of our recent blog entries). Therefore, speech provides a large amount of diagnostically valuable information. Depending on the appearance and symptomatology of a disease, divergent speech patterns emerge that make it possible to distinguish between different diseases and their progression. The term “speech biomarkers” refers to a group of altered speech properties—so-called features— that are characteristic for a particular disease or symptom.
How do we derive specific biomarkers?
In order to combine relevant speech features into a concrete speech biomarker, a large amount of symptomatic speech data is needed. This data can be processed on a linguistic and a paralinguistic level. The linguistic analysis covers “what” was said and the paralinguistic analysis covers the evaluation of “how” something is said (e.g. loudness or melody).
On a symptom level, the content of the language, i.e. what was said, is often more associated with cognitive abilities while how something is said is more associated with mood or motor ability.
This of course is just a rough grouping that neglects the context-dependency of language. For most neuropsychological assessments, context means the task given to the patient. Free speech is rarely used; instead, specific, standardized constraints are often imposed. For example, subjects must tell an emotional story or name as many animals as possible in a short time or describe a picture. Semantics (the content of language) as a linguistic feature can be interpreted differently in the two tasks. In the animal naming task, semantic relatedness of individual animals implies a workingstream of thought and would thus approximate cognitive abilities. In the case of telling an emotional story, a lot of speech or little speech could be produced depending on valence, so semantics would be an approximation for mood here. Thus, depending on the context, speech features can be predictive of different types of symptoms.
Speech analysis is useful at different timepoints in a patient journey. It can be used as screening, for specification of a diagnosis and also for disease monitoring. Since speech analysis is non-invasive and can be used remotely, it offers opportunities for assessment with high frequency. In addition, speech analysis can also be used in clinical trials, for example to support a labeling claim for a new drug or to recruit an appropriate sample for a trial. For this purpose, ki elements offers support during the complete study. ki elements offers support in identifying a use case, consulting on technical feasibility, analysis of the data and help with licensing for validated biomarkers and ML models for new products.