THERE IS A REALISTIC CHANCE WE CAN DETECT COVID-19 FROM THE VOICE
- COVID-19 is a lung disease with symptoms that affect speech, e.g. dry cough, unusual way of catching breath while speaking, irregular breathing patterns.
- Medical voice sound analytics has made big progress in the last decades - it was shown multiple diseases like Parkinson’s, Depression or Alzheimer’s are detectable from voice sound.
- We are a leading player in voice analytics, with >20 years of experience and funding of a Horizon 2020 programme. We just finished one of the largest clinical trials in the field ever together with Charité Berlin.
IN CASE OF SUCCESS, THIS WOULD BE A REAL BREAKTHROUGH AGAINST THE DISEASE
- Distinguishing COVID-19 from normal flu early would allow systematic telephone screening & isolating of patients - which might reduce cases and help preventing a new wave.
- It could help monitoring the disease at home and provide more insights about the symptoms to the global medical community.
- Potentially, it might even be possible to predict severe from non-severe courses of the disease to allow for optimal treatment.
SEVERAL COMMERCIAL PLAYERS AND ACADEMIA ARE TRYING TO SOLVE THE PROBLEM USING VOICE DATA FROM THE PUBLIC
- Several commercial teams are trying to solve this problem by asking the public to donate voice data.
- Also academic institutions are trying the same, see for example the University of Cambridge.
- For what we see from the outside, especially publications from academics mostly use very simple technical approaches - either applying any sorts of neural nets on the raw data, or by using simple open source sound features from free toolboxes.
THESE APPROACHES WILL FACE MASSIVE TECHNICAL, MEDICAL AND REGULATORY CHALLENGES ONCE THEY TRY TO BE USED IN PRACTICE
- Technical problem: with the simple approaches above achieving high success rates is still possible - however the algorithms are often not stable when used with real data unless they were built on a very large amount of well-labeled training data - something that is extremely difficult to receive.
- Medical problem: since no doctor is involved in the training data collection, getting well-labeled data is at least highly doubtful given all the medical complexity (e.g. different phases of the disease, symptomatic vs. asymptomatic, COVID-19 vs. condition from other Corona viruses, etc.).
- Regulatory problem: at the end, even if a stable solution could be built that covered all the medical complexity, it would not be eligible for medical certification since no verified diagnosis was involved. We do not intend to make a legal statement here, and emergency-regulation might be different - yet making a diagnostic claim, even if you call it “pre-diagnosis” or “pre-screening” without approval as medical device is typically very far away from legally safe.
WE ARE THEREFORE TAKING A RIGID ROUTE AND BUILD ALGORITHMS BASED ON CLINICAL DATA AND FEATURE-ENGINEERED ALGORITHMS
- We are therefore deliberately taking a clinical route. We believe it is the same as with vaccinations and drugs - a rigid, hard-medical route may be slower and less “catchy”, but much needed at the end. Thus, we will only work with voice data from properly diagnosed patients.
- On that basis, combined with our technology, which has been built over more than 20 years and is highly feature-engineered - we can resolve all problems mentioned above and potentially create a voice solution to really work in practice.
DURING THIS HACKATHON WE'VE
- widened our network of key stakeholders by establishing contacts to clinicians, big pharma, clinical research professionals
- advanced our clinical research plan
- developed a save and data-secure data-acquisition concept applicable in the clinic.
SPECIAL THANKS TO ALL OUR MENTORS FOR THE AWESOME DISCUSSIONS