one of the founders (Tanay) here - happy to answer any questions. we also put out a new trove of data here: https://athelas.com/publications
* How do you account for interstitial fluid expressed into the droplet?
* Do you use the same flags as a CBC analyzer?
* Have you tested blasts?
* Have you tested leukostatic samples (very high white count: over 100k)?
* Have you validated against platelet clumping?
* Have you validated for basic sources of interference (lipemia, hemolysis, icterus).
* Have you engaged Cerner? Epic? DrChrono? AllScripts? Other EMR providers?
* Have you engaged a pathologist?
I think this device is super awesome and I was trying to come up with something similar a few years back.
The primary concerns is that one of the first things I learnt in Med School is that capillary blood is way less reliable in compare to venous blood for something as important as CBC.
Not sure how they are accounting for the issue that the sample they have isn't the best one to get started with.
ps: killjoywashere how do I get in touch with you ? Am at Asingh [at] healthenclave.com
-The first line of defense against interstitial fluid is wiping away the first drop (commonly used protocol in most drop test). The inevitable remaining interstitial fluid chunk is then classified out by our image processing models, which have been trained to recognize differences in images based on debris/cell concentration.
-We currently report the WBC, WBC Differential, Platelets, and RBC indices. The flags we're missing are secondary RBC metrics like MCH and RDW estimates but we have plans to roll these out soon as well.
-Yes, we've run both bench and clinical testing on patients with north of 100k white blood cells - the image processing can segment out cell chunks in these crowded samples better than traditional flow cytometry. Not in the currently published studies, but some good data coming out about this soon.
-We've run basic initial tests on platelet clumping (artificially induced in certain diluted samples), and present in clinical samples. Since our model looks at the boundaries of these platelet cells, we have a lot less trouble with this than flow cytometry/beckman and even human pathologists. Still, we're working to find more samples with clumping to define the limits of detection on this front.
-We've done interference studies with Hemolysis, EDTA, not with lipemia and icetrus. These are on our list of clinical samples to source as well.
-We've begun working with some of these EMR providers, the good thing is many of them have sandboxes to get setup in relatively quickly. The larger ones do have long, long-term engagement timelines.
-We work with dozens of pathologists to help review results, train the system further, and in general go over good morphology practice. Has definitely been a core part of our strategy.
What are the details behind $20/month? E.g. is the machine "rented"? Or is the $20 for consumables?
Is it connected to the cloud? (In other words, is the machine-learning algorithm's "counting" happening in the device or are you uploading image data to a server for analysis?)
Do you control for age and other bio-factors and how?