Kintsugi Health and Healthie

The Kintsugi and Healthie integration offers healthcare organizations a solution to aid in the early identification and treatment of Depression and Anxiety, boosting cost-savings and outcomes for patients. Kintsugi’s solution analyzes the presence of signals consistent with mental health episodes in patients from short clips of free-form speech captured during your patient encounter within Healthie. 

Note: This integration and API solution is available to all Healthie healthcare organizations on Plus, Group, and Enterprise plans, with the creation of a Kintsugi account. 

Create a Kintsugi Account

The Kintsugi API uses a unique API key that is provisioned specifically for your organization to authenticate your API requests. KV surfaces an invalid request error when you do not include an API key and an authentication error when the key is invalid.  

To be provisioned with an API key, Kintsugi customers should send an email to engineering@kintsugihealth.com with the name of your organization, the email that needs access, and mention that you are a Healthie customer. Once approved, we will allowlist your email and send an activation token for your organization.

Enable Kintsugi Integration with Healthie

Once you’ve enabled a Kintsugi account, email hello@gethealthie.com for your API Key. The Healthie Support Engineering team will share the API Key to your Documents section which you can share with Kintsugi.

Fees (from Kintsugi)

Kintsugi Health is available for Healthie customers on Plus, Group, and Enterprise plans; a Kintsugi contract is also required.

Whether you want to request a demo or a press kit, are in clinical practice and want to help with research, or just have more questions about Kintsugi, please email support@kintsugihealth.com.

Additional Resources 

• Clinical Validation Studies on Kintsugi’s Voice Biomarker Technology
• Kintsugi Blog highlighting case studies and customer deployments
• Kintsugi offers two models for depression identification: Severity Depression Model vs Binary Depression Model
  • Binary depression model: the binary prediction detects the presence or absence of current symptoms consistent with depression and returns a true or false result.
  • Severity depression model: the severity prediction stratifies the magnitude of the patient's current depressive symptoms into one of three categories: No to Mild, Mild to Moderate, and Moderate to Severe.
• [Coming soon!] Healthie + Kintsugi Health Integration flow - patient consent, audio capture, prediction return, results view, post-visit result storage

Kintsugi FAQ's

How does the technology work?

• Kintsugi is developing voice biomarker software to detect signs of clinical depression and anxiety from short clips of free-form speech.
• Kintsugi uses a combination of deep learning and signal processing techniques to understand the common patterns associated with clinical depression and anxiety which are different from temporary feelings of sadness or anger, etc.

How were your models validated? Trained?

• In the early part of Kintsugi, we developed a consumer voice journaling application to solve the challenge of not being able to always speak with a therapist and having self-tools to alleviate stress, anxiety, and depression. The consumer-facing app has over 200K downloads and users in over 250 international cities which has given us a data set hundreds of orders larger than the largest publicly available DAIC-WOZ dataset.
• We trained our models on a large proprietary dataset with self-annotated labels of depression and anxiety. In addition, we capture additional datasets through our clinical, academic, and commercial works in market through IRB-approved studies. We validated our model using mental health professionals with 7 years+ experience and compared the binary results of their diagnosis with our model’s performance. We are nearly 2x as performant as primary care doctors who today administer 70% of antidepressant scripts.

Why does the technology work?

• We use 100 different muscles in our chest, neck, and throat to push the sound out of our mouth as voice
• Voice biomarkers have been studied for almost 100 years starting with Emil Krapelin’s work that identified certain features of the voice correlated with depression
• In the last decade, the acceleration of deep learning models and the availability and cost of compute has allowed us to mathematically model patterns in large data sets
• Humans are able to hear up to 20 kHz and in telehealth calls, these are held in voice-over IP calls where the transmission of audio fidelity is 44.1 kHz, more than double the data per every second.
• For therapists who have had a few hundred clients over their careers, a machine-learning model can ingest tens of thousands of labeled examples to develop a mathematical representation of what depression and anxiety look like
• Classical features like pitch, prosody, slowness of thought (psychomotor retardation and agitation) have shown to be evident in depressed and anxious speaking across East and West clinical studies

Can you trick the system?

• You can trick the system, but the question is when you are having a regular conversation with a doctor, why would you speak differently?
• It is important to note that our models are looking at underlying characteristics of depression as evidenced by physiological characteristics which are different from temporary feelings of sadness, anger, happiness, etc.