I spent a long time discussing Snaq’s use of AI for food recognition in the previous article. In this one, I’m looking at what it does in relation to glucose patterns.
Snaq looks at two areas:
- Feedback on your time in range for the three hours after a meal;
- Providing an estimate, given the information it has, on what will happen to your glucose levels following a meal.
For full disclosure, I was using Snaq alongside Boost, so I entered no data relating to mealtime insulin dosing. Having said that, I also couldn’t find a way to do that, so it may just be their marketing that contains the dosing information, due to the app having no medical approvals.
I assume the model is capable of taking the insulin data onboard, and with appropriate user settings, making recommendations about mealtime bolus adjustments based on historic data, but it’s not something I saw during my use.
At this point, I’ll also highlight that for the predictive capabilities to work fully you need at least 5 meals and three days of data.
So how does this all work?
Predictions
This is as straightforward as it sounds. The model in Snaq observes what happens over the previous 3 days then starts to offer up a suggestion of what might happen based on what it observes.
As can be seen from this example, the estimate gives a range for potential rise and a similar range for potential time to peak, which you can then track in the outcomes pane once the meal is deemed complete.
In this additional example, we have a much higher carb meal.
Now, the question might arise, why, for a much larger carb load, is a similar rise and peak duration estimated. Of course, if your ratios are nailed, it should make very little difference, or in this case, if Boost is working well for you, it equally should make very little difference.
And that’s what we’re seeing in these two examples.
As ever, there are certain scenarios where I have questions. The main one being the overlap of two “meals”.
On a hot Sunday morning, I had breakfast, then around two hours later, I was hot and had an iced latte.
All well and good, but when we take a look at the results of these two meals, we see something interesting.
The peak at 2h16 after the first meal could easily have come from the latte, especially as the glucose levels appeared to have started to fall when that was entered.
There is another example of compounded meal entries, by way of Chicken Enchiladas, followed by a mini magnum and M&S chocolate coated custard cream.
The peak for all three entries is given as the same maximum number, however, it’s unlikely that each on their own got to that level. Instead, it’s the combination that achieved that, and Snaq is unable to distinguish between the attribution of each entry in that process.
Post-meal feedback
When the time post your meal is up, it sends you a notification, much like the set below, providing feedback.
When you click through these, you get the data used in the first section describing time in range, peak, time to peak and rise.
Firstly, the positive feedback when the outcomes are good is pleasant. It’s reassuring to know that you’re bolusing well or that your bolusing algorithm is working well.
The second part of this relates more to how you (or the algorithm) calculated your dose, and whether there are things you could change.
In all the examples here, post-meal we can review whether the timing and size of our bolusing is working for the type of meal we’ve eaten.
For example, if, after a meal, I had a low time in range with a larger climb, later peak and more time above range, I might conclude that I needed a larger mealtime bolus, and/or that I should take it earlier before the meal. Likewise, if I end up with less of a climb, lower TIR and significant time below range, I may look at dosing later or perhaps dosing less.
It also has the potential to help you to choose a square wave bolus for prolonged high type meals, and given the meal tracking, peak data, etc, make it easier to choose an appropriate duration.
Back to the £100 question. Would I pay for it?
For those who are TL:DR, no. I wouldn’t.
It doesn’t offer me enough utility to make it worth the extra cash over a year of carbs and cals at £35, for example.
The amount of work I felt was required to correct the meal recognition AI (which was for everything I entered other than the breakfast plates), means that meal entry becomes a chore. And other apps out there, such as Under My Fork allow you to track the meal and glucose response, without having to train someone else’s model, with the same lookup functionality.
Snaq feels like a system looking for a buyer. I get the impression that the model probably already has features that would allow it to make recommendations around insulin dosing and adjustments. In order to fine tune some of the predictive components and obviously, due to the cost of trials that would be required to get approvals for such behaviour, this isn’t available, and instead we have what feels like a public beta.
The big question with start-ups like this is “does it have what people want?”. More than 50,000 people have downloaded the app from the play store. I dare say a similar number have from the App store. I’d love to know the stats for retention though, because that feels like the biggest challenge.
When we talk about new diabetes apps, we always need to consider utility, and for me there’s just not enough. It is simply too unwieldy. However, I’m probably a minority user.
There’s another aspect of this to consider. Where are future treatments going? I’ve deliberately mentioned Boost throughout this article because its whole purpose is to eliminate the need for an app like this.
And I’m not the only one that thinks this way. Medtronic, Control-IQ and CamAPS have all discussed and have been testing approaches to handling meals with no carb entries.
While these are (mostly) for the future, they highlight the direction of travel, and while the majority of the the global population only has access to injection based insulin at the moment, I’d like to think they’ll gain access to better automated solutions in the not too distant future.
So while AI is hot, and while people need it, I’m sure that Snaq can make a living from selling its subscription model to users. I just won’t be someone paying.
Leave a Reply