Project / product name: CKD Watch
Link to the project: https://github.com/DataSentics/EHH_DATAMASH
Team leader: Tereza Novotna
Challenge: 6. Are you kidneying?
Problem: 10% of people has CKD, great part of them find out late, causing them to be stuck on dialysis for life. This could be easily prevented if more GPs send their patients to regular lab testing. These tests are not expensive and can save both life quality of the patient and a lot of money/time to healthcare system. The interest is in exploration if there are any other factors which might be used to predict CKD and build reusable algorithm which counts with the fact that the data are very sparsed.
Solution: We developed solution which connects data from labs and personal info and train ML model on them to predict lab test results confirming disease. We can then get the best predictors and visualize results. Model can natively deal with sparse data, part of the solution being also feature selection. Next step would be auto-generate report, send alert to GP in form of short notification as well as link to the full personalized patients report. Another option would be also alert patients themselves.
Impact: The CKD is highly prevalent disease – about 1/10 patients is susceptible to CKD. However, it often goes undiagnosed, especially in early stages, leading to extensive costs related to RRT. It amounts to about 6 billion of CZK yearly in Czech Republic only. The critical GFR tests are often missing. The CKD Watch solution can alert and raise awareness of the GPs for CKD risk of their patients, based on other available lab tests, risk factors and thus help to mitigate significant share of the costs.
Feasibility & financials: The solution is designed to run with limited amount of lab tests and other data available. CKD watch could be integrated with the GPs usual application to run the analysis for his patients. Using DASTA or FHIR is neccessary to make this solution to be applicable. Financial estimate for 6 months with 2.5FTEs of combined capacity data scientist, engineers and fullstack developers would result in approx. 140 000 Eur.
What is new about your solution?: Our solution keeps in mind this problem is highly caused with "human factor" and negligence or simply general unawareness of possible benefits of early preventions. Key idea is to be able to notify doctors about patients in risk and educate. It has two parts, one based on ML one based on commonly known correlated conditions with CKD which are always part of alert if they are in patients anamnesis regardless of ML. Connecting trusted experience with AI leads to trust in solution.
What you have built at the hackathon - text explanation + code (e.g. GitHub link): CKD watch is designed for alerting GPs in regard to CKD combining known risk factors and predictions from machine learning model of eGFR test results, that we trained on alternative lab tests and patient characteristics. https://github.com/DataSentics/EHH_DATAMASH
What you had before the hackathon, please mention open source as well: We had nothing prepared in advance.
What comes next and what you wish to achieve: Solution was build with further development in mind. Further development would require discussion with medical experts as well as med. software developers and/or hospitals. Also further discussion about options to get results directly to patient should take place. Model is build to be flexible and should not lead computational overhead. Adding standardized data protocol should be easy.