Amazon announced early access for customers in the US to its first health gadget, Amazon Halo, a combination of an app and band, that claims to recognise emotion in a user’s voice using machine learning, and calculate body fat percentage using machine learning and computer vision. All health data is encrypted in transit and in cloud, and can be deleted by the user at any time. When it is stored on the phone, it uses the phone’s “full disc encryption”.

It is currently available for a reduced price $64.99 in the US against a list price of $99.99. During this early access period, users will also get 6 months of Halo membership that will renew automatically for $3.99/month after the first 6 months. Non-members have access to only basic features such as step count, sleep time, and heart rate.

What data do the app and the band collect?

  • The Halo Band measures: skin temperature, motion, heart rate.
  • The app collects information about: fitness metrics, body fat composition, demographic data, sleep, tone of voice.
  • Apart from this, the Halo service, as a combination of the Band and app collects data on how users use the service through metrics such as: if the band is on the wrist, plugged into a charger, how often particular pages are opened in the app, feature usage, etc.

How does the voice-based emotion recognition work?

Amazon calls its voice-based emotion recognition system Tone. It uses “machine learning (ML) to analyze the positivity and energy of your voice” to see how you sound to others. The Halo band and app use voice detection algorithms to “pick up speech, remove background noise, and optimize battery life”. AI then analyses the qualities of the customer’s voice “such as pitch, intensity, tempo, and rhythm to predict how others would perceive and describe the customer’s tone of voice, which creates a summary you can see and use to identify trends within your life”.

Privacy features: Amazon has said that speech samples, that are used for Tone are always analysed locally on the user’s phone and automatically deleted after processing so that nobody, including the user, ever hears them; they can’t ever be played back. They are never sent to the cloud. Both the band (through its two microphones) and the phone app are used to collect speech samples, depending on the features the user is using. The microphones on the band can be muted. Recording on the band are automatically deleted within three days if the band is not connected to the app (via Bluetooth). When recordings are transferred from band to phone for processing, they are deleted from the band.

How will Halo calculate body fat percentage?

Through its new feature, called “Body”, Amazon will use computer vision (CV) and machine learning to measure and track a user’s body fat percentage (BFP) using the phone’s camera and the Halo app. Using CV and ML, the algorithms create a 3D body model of the user, and calculate the BFP. It also has a “body model slider” that shows how a user’s body could change as they gained or lost fat.

Its accuracy is limited for users with more than 50% BFP, those who are over 500 pounds (226 kilograms), pregnant women, those who use wheelchairs, and people with certain physical differences such as missing or prosthetic limbs”, the product page says. Amazon has offered no disclosures about the accuracy rates of its sensors or algorithms that will do all the processing

Privacy features: Users can download or delete their data at any time from the app. Body scan images are automatically deleted from the cloud after processing so that only the user sees them. The images and the 3D body model are only stored locally on the phone. It is the user’s choice to opt-in to store their Body scan images in Amazon cloud.  

Third party linking is possible

The app allows users to link their Halo accounts with third party programmes through “explicit opt-in” by the user. Amazon Halo has been integrated into Cerner’s products. Cerner is a Kansas City-headquartered company that maintains electronic health records and provides other health IT services. This will supposedly allow users to “share health information with their care teams and directly into their electronic health record (EHR)”. This way, doctors and hospitals that are Cerner clients would directly get access Halo users’ health data collected by Halo. It is not clear if users will have a say over what kind of data, both from sensors in the band and data produced by ML algorithms, is integrated into their EHRs. AWS is already Cerner’s preferred cloud partner.

Amazon already offers this for members of WW, formerly Weight Watchers, and the John Hancock Vitality wellness program. Other partnerships will appear in the Halo app as “labs” and these include Mayo Clinic, Exhale, Aaptiv, Lifesum, Headspace, and others, as per the Verge. Amazon has said that it will audit these labs and data created from the fitness challenges that Labs create for users will be shared with the partners only in an anonymous, aggregated manner.

The emotion recognition feature is terrifying, as are Amazon’s network effects: MediaNama’s take

On emotion recognition: Simply put, the emotion recognition feature via speech is terrifying. Amazon’s earlier foray into facial recognition, via Rekognition, was proven by ACLU to be discriminatory and more inaccurate for people of colour, especially African Americans. A December 2019 report from the National Institute of Standards and Technology (NIST) showed that of the 189 facial recognition algorithms that were submitted, many were 10 to 100 times more likely to misidentify a black or East Asian face compared to a white face; accuracy rates were the worst for black women.

A May 2020 academic paper showed how popular automated speech recognition systems such as Alexa, Siri, Google, Watson and Cortana, fared worse in transcribing speech from black speakers. Amazon’s Tone analyses a user’s “pitch, intensity, tempo, and rhythm to predict how others would perceive and describe the customer’s tone of voice, which creates a summary you can see and use to identify trends within your life”. All these factors are culturally determined, and their perception is similarly subject to cultural biases centered around stereotypes. Amazon itself acknowledges that Tone “currently works best for American English speakers”, assuming that American English is one homogenous whole or that there is only one American English accent. The biggest casualty in this case, of course, could be African-American Vernacular English (AAVE).

Pricing model suggests focus on revenue from membership: The pricing model suggests that Amazon is banking on revenue from the membership and allied services rather than device sales itself. The membership service will become more lucrative for the company only if it monetises the data through partnerships and integrations with third parties. Given the wealth of data Amazon has on its users/customers, when health data is cross-referenced against people’s buying habits on Amazon (especially Pantry, Whole Foods, Amazon Fresh, etc.), it can be used to create health profiles of people and probability of risk for different diseases.

Health insurance companies can use this to discriminate against people with pre-existing conditions, and those at risk of developing certain conditions because of the lifestyle choices their socio-economic status forces them to make. Since race and class tend to coincide in the US because of systemic and historical structures of oppression, people of colour, and working classes, would be further locked out of the debilitatingly expensive American healthcare system using such data.