The first cohort of Amii’s Machine Learning Technician I Certification Course has completed their course work and presented capstone projects to a panel of AI industry experts including scientists, researchers and business experts. Students in the first cohort included representatives from Amii’s industry partners, spanning industries such as government, oil & gas and higher education, as well as individual learners.
As a final project in the course, participants were required to scope out a specific problem using COVID-19 datasets from the UNCOVER COVID-19 Challenge on Kaggle.com hosted by the Roche Data Science Coalition. Students implemented three supervised learning algorithms – two for classification and one for regression – to create Question-Answering Machines (QuAMs) that support business decision-making at a theoretical company. In teams of four, students scoped their problem, cleaned, analyzed and visualized data, built a QuAM and then evaluated the effectiveness of those models from both a technical and business perspective.
Predicting risk-levels by geography
The first group sought to reduce the number of COVID-19 fatalities using US census data from states and counties and daily reported number of deaths. Given mortality trends, students determined different levels of risk of diverse regions from week to week, thus gaining geographic insight into the spread of the virus. Ultimately, machine learning could help guide policymakers’ decision making in prioritizing the delivery of resources and healthcare personnel and enacting or removing social distancing orders and travel bans.
Using blood tests for diagnosis and treatment planning
The second group aimed to contain the spread of COVID-19 by accurately diagnosing patients using variables extracted from standard blood tests. Using a combination of technical and domain knowledge, the group offered a comprehensive view of the problem that considered the consequences of providing front-line healthcare workers with a machine learning model as a decision-making tool for diagnosis and treatment planning. One important finding from this group’s proof-of-concept is the potential for significant reductions in the number of test kits needed, especially in areas where test kits are very limited or completely inaccessible.
Predicting pandemic vulnerability
The third group focused on a specific population, with the goal of identifying which areas would be most susceptible to COVID-19, based on socio-economic factors. The group identified the potential of using machine learning to rank levels of vulnerability of different areas based on current understandings of how pandemics impact vulnerable communities. The group anticipates that their approach could help reduce the unequal impact of COVID-19 or similar infectious diseases in vulnerable and at-risk communities.
Hospital admittance prediction
The fourth group used data from hospital admittance to predict which positively-diagnosed COVID-19 patients would have needed hospitalization. This would help healthcare workers identify which patients should be immediately admitted, leading to more effective treatment and successful recovery of patients. Due to the sparsity of data available and the missing information about patients admittance, the group identified critical aspects of admittance that may impact patient recovery. Factors included the variability in admittance versus hospital capacity, as well as the variability in the available resources that may result in patients being turned away due to a lack of beds, ventilators, or staff.