Author:
CDT John Murray
Harvard College
Co-Authors:
John Murray (Harvard), Liam Conboy (MIT), Jack Cogbill (Harvard), Grace Motes (Harvard)

Abstract

Wearable devices are extremely prevalent these days and they collect a myriad of data. However, what good is this biometric data if we don't understand what it means or if this data is not presented to us in an informative way? To tackle this issue, our team has focused on creating a program that would take a few of these data points that are collected by a Garmin watch in conjunction with a soldier's baseline body metrics (weight, gender, etc..) to provide an informative summary of their performance and readiness. Specifically, we have first decided to create a program that identifies a soldier's relative risk to heat injury based on a suggestion by USASOC at the MIT ROTC Hackathon weekend back in October.

Problem Statement

Since many operators and soldiers from USASOC are training and deploying at such a high tempo, longevity and readiness are crucial. We see our program as an essential tool to help aid soldiers and operators as they prepare for their various missions and tasks. Heat injury is continually stressed throughout the Army top to bottom, so we wanted to create a program that would identify a soldier or operator's relative risk to heat injury based on how they are performing before it even occurs. Our program will help keep people in the fight and will provide an informative narrative based on loads of data that is collected by the watch that many operators and soldiers are already wearing. Additionally, we see our initial Heat Injury relative risk program as a proof of concept for many other programs that can be derived using various biometric data points, displaying reasoning for a more robust wearable device program.

Proposal

The Soldier Health Metric Dashboard is a smartphone app that allows an organization to track the health information of all of the organization's soldiers in real-time. Soldiers can wear devices, like smart watches, that keep track of their physical health data. That physical health data is fed into an algorithm that calculates the injury risk level of each individual soldier. The Soldier Health Metric Dashboard app then reads the injury risk score calculation for each individual soldier and displays that data in an easy-to-understand way. The end-users of the app (Platoon leaders, Squad leaders, etc...) can then receive alerts and notifications to their smartphone devices when a given soldier's heat injury risk score is above the acceptable level, and they can then use the app to view that soldier's health data and see the reason why their injury risk score is above the acceptable level.

Challenges and Unknowns

Our biggest challenge/unknown is the accuracy of the algorithm since we have not been provided with any data sets so far. If we were given data sets, we would be able to use them to fine-tune our algorithm so that the heat injury relative risk is as informative as possible. 

Comments

kcole | 8 February 2021

There is an obvious parallel here to firefighter rehabilitation on a fire scene.  To my experience, this is all done with manual intervention by EMTs on the fire ground.  In Massachusetts (and likely elsewhere) a two-bottle rule is applied after which periodic rehab occurs and stable vital signs are required for re-entry - however firefighters have been known to avoid rehab stations to stay in the fight and cardiac events are still one of the leading causes if line-of-duty deaths in the fire service.  A device like this could have benefits beyond soldier health to stressful occupations.

DFreinberg | 9 February 2021

I agree with "Kcole" the applicability here is certainly larger than the special operations community - think athletes and anyone in the uniformed services too.   

  • What elements are you monitoring?
  • What are the warning signs of a heat injury?
  • Should the system (watch?) notify the soldier (and leader?) of impending injury and guide the Soldier through a de-escalation?
  • How should the data be protected?
  • Why Garmin? (versus Apple Watch and others)
  • How does this work in an austere area with no connectivity?

dBlocher | 9 February 2021

I agree with your assessment that a big challenge would be the accuracy of the algorithm. It might also be more challenging than that - A fundamental question that DFEINBERG mentioned is what are the right signals to monitor - is it just body temperature over time? What about fluid intake?

atindall | 20 February 2021

Awesome idea and definitely how we need to think in reference to soldier performance. Some thoughts after reading the proposal and comments, feel free to reach out to chat more.

  1. Getting the Data
    1. Think there are a couple options here for data to train an algo – in my opinion quickest route to de-identified data is reaching out to 75th Ranger Regiment RASP Program or the Q-Course where heat injuries are tracked closely.  Note sure how many bio-markers they actually track though that would help predict when to intervene – potentially we could look at a trend of degraded performance on physical events plus environmental decisions to intervene. They definitely track a time series data everytime they run the program.  Also, there was a team at MIT LL under Dr. Reed Hyot in the USREM group that had built a core temp sensing device for the 75th’s RASP program – maybe a way to get some actual operator core temp data.  I think this could also be a good “customer” engagement to think about how to design the system with the user in mind – probably will have to do testing in a “training” environment so may be good to make a connection to one of these programs. Happy to help with reaching out.
  2. UI Design & SOP Integration
    1. Training programs have an extremely strict process for assessing heat injuries. This is a great place to start – I would fully understand this decision cycle and how your product enables an instructor, and in the future a leader on the battlefield, make an informed decision about when to take action. Potentially, there are user inputs into the interface to inform the algorithm’s recommendation – one key marker is “disorientation,” this is an extremely subjective data point, experienced leader’s understand it, new ones maybe not as much—could we do this with computer vision or speech recognition on a smart phone at the edge?, looking at pupil dilation or operator gait or slurred speech. Without getting to technical, suggestion is think about how we can quantify a historically subjective assessment relying on human intuition …
  3. Hardware/Software Integration
    1. This is obviously a concern at the edge. Think about some new edge computing devices: here is a good use case that could stimulate your thoughts https://community.arm.com/developer/ip-products/processors/b/ml-ip-blog/posts/train-a-tinyml-model-to-recognize-sounds-that-uses-23-kb-of-ram?_ga=2.105120330.810538440.1612009146-1137999121.1612009146

 

https://www.arm.com/campaigns/arm-tinyml?utm_source=linkedin&utm_medium=social-paid&utm_campaign=2020_ai-ml-ecosystem_mk09-5_1000heads_social-prospecting_linkedin_na&utm_term=algorithm-provider&utm_content=image-tinymlimage-developer&li_fat_id=896714e3-8074-48f6-a6ad-40bd37974675

 

And some relatively cheap NVIDIA devices:

https://developer.nvidia.com/embedded/jetson-nano-developer-kit

 

Again, happy to chat more about these

 

  1. What biomarkers matter?
    1. Obviously, a thermocouple sensor to measure surface temp would provide one data point. Might be good to get collaboration from a medical provider here – I am by no means an expert but I know different people will run at different surface temps and even core temps – how to you baseline the marker to the user? Can we actually make a prediction from skin temp.  
    2. Other biomarkers to think about – there have been huge advances in blood glucose measure devices – again, need a medical provider input here but is there any correlation with low blood glucose and dropping from heat. I know lots of times soldiers need to eat but, maybe affects electrolyte levels more than blood glucose. There are a couple companies using sweat to measure hydration level (nix biosensors) (https://nixbiosensors.com/) … might be worth looking into one of these Commercial of the Shelf solutions and integrating into your system. Not sure this one has the capability of transferring the data but could it be modified to fit into your system?
    3. Another area that is expanding is eSMG enabled compression clothing to provide muscle activation signals – could be another good data point to think about in a training environment. One company that is doing this is athos, https://shop.liveathos.com/ … they have contracts with numerous pro sports teams and are looking to break into the defense sector. Right now the use case is primarily focused on proper lifting form to prevent lower extremity injuries … however, they have a method for capture exo-skel data, the algorithm could be shifted to think about how muscle firing correlates in the case of heat injury. They also capture heart rate data, don’t think they have temperature but may be worth connecting to discuss. They use a simple iOS “coaching” app like you guys are talking about developing so could be an interesting integration.  They are trying to do this "risk" score concept. 
    4. Another pretty good dashboard that you could try to integrate with is the commander readiness dashboard -- actually one of the better products the army has fielded on the software side in my opinion. Challenge here is it pulls direct from the medical side so I think security and access would be an issue. 
    5. As you probably know, More features = more complexity != a better model necessarily, so I would encourage a lot of thought about which metrics we think actually matter.

 

 

Hope this helps stimulate some design thoughts, would love to connect and chat more about some other challenges I think are on the horizon here – feel free to reach out at atindall@mit.edu