Introducing Kino v2.1

Kino v2.1 is our most capable model to date. It shows the ability to generalize across body types and demographics with accuracy unlike any version before it.

v2.1 is our first model that experienced tangible gains from data-scaling - leveraging our breakthrough architecture from our v2 class models. This tangible boost in performance can be attributed to an influx of professionally collected data that spans a broad demographic set.

With additional minor improvements to Kino's embedding extraction, Kino v2.1 handily surpassed previous visual body composition (VBC) benchmarks, while also outperforming the InBody 570 in our internal tests covering a population-representative, held-out test-set.

Kino v2.1's capabilities

All tests in the following section compare Kino's current and previous 'production' endpoints to the Hologic DEXA's tissue body fat % prediction. Predictions were derived from a held-out test set of 195 individuals spanning 7-60% body fat, 18-75 years of age, and a representative ethnicity and gender split.

What's improved

v2.1's most notable improvements include the addition of three additional body composition metrics that surpassed our internal accuracy threshold for production (R² ≥ ~0.800), as well as increased accuracy for diverse user populations such as heavier users and women.

Foremost, v2.1 is the first Kino model to surpass internal accuracy thresholds for lean muscle mass estimations made directly from photos (without the aid of body fat estimation as a prior). v2.1 also showed convergence for appendicular lean mass, as well as segmental lean mass for both arms and legs. More specifically, v2.1 showed the ability to estimate lean mass % in participants with a coefficient of determination of 0.842 — outperforming all prior models in this task.

Additionally, v2.1 saw tangible improvements among more diverse participants, which we attribute heavily to an increase in diverse training data. For example, mean absolute error (MAE) for female users dropped from 3.42 percentage points to 3.09 percentage points from v2 to v2.1. This represents a 9.7% relative improvement in body fat estimation for women.

v2.1 recorded the lowest error of any Kino model to date, clocking an MAE on our held-out test-set of 2.77 percentage points. This marks a significant improvement from our previous model, v2, which recorded an MAE of 3.41 percentage points on the same test-set.

Looking across body composition estimation methods, Kino v2.1 also outperformed the InBody 570 in our internal tests when both methods were sampled over a test set of 195 unseen participants. v2.1 recorded roughly 1.7 less percentage points error than the InBody 570 (38% less relative error) in these internal tests when comparing body fat estimates to the Hologic DEXA's tissue body fat %.

v2.1 is Kino's first model that showed the ability to estimate lean mass % and appendicular lean mass from image-data only. This marks a true phase-shift in the model's capabilities as we are able to derive strong enough signal from visual data alone such that the model's embeddings were sufficiently informative to estimate lean mass % and appendicular lean mass from these embeddings alone.

Specifically, v2.1 showed the ability to "explain away" 84.2% and 79.8% of the variance when estimating muscle mass % and appendicular lean mass respectively — solely from image data.

Quantitatively, v2.1's muscle related outputs still lag our body fat % accuracy. However, we are internally forecasting improved accuracy on muscle mass metrics (namely lean mass % and appendicular lean mass) in subsequent models. At the current pace of improvement, we expect our models' lean muscle mass % estimations to outperform body fat % estimations within the v2 class of models.

Also launching today

Along with Kino v2.1, we are also launching a client check-in workspace for our fitness partners.

Kino's check-in workspace is available on the Kino Fitness Dashboard, and allows physicians, personal trainers, and other health and fitness professionals to show physical progress to clients and patients alike.

With the check-in workspace, you can build deeper commitment to care plans with your patients and clients through data-backed progress reports — proving what you are doing is making a difference.

The check-in workspace allows professionals to log notes and correlate them with changes in body composition. For example, if a patient begins using a GLP-1 or other supplement, like creatine, the care provider can now visualize the results of these interventions, while coordinating with our HIPAA-compliant and in-platform chatbot to further strategize patient care plans.

Availability

As of June 7, Kino v2.1 is fully available to all existing Kino partners at no extra cost. To date, we have not introduced usage limits so members are free to scan as frequently as they like.

To gain access to state-of-the-art visual body composition models like Kino v2.1, reach out to us at jacksongerard@kino-fitness.com