Data Skeptic: Haywire Algorithms
From Data Skeptic:
Today, we are joined by Ravi Parikh, a Medical Oncologist and an Assistant Professor at the University of Pennsylvania. He runs a lab that develops and implements machine learning predictive models in clinical care. Ravi discusses his research on how the pandemic has toppled the performance of machine learning models in the medical field.
The medical researcher kicked off by establishing how he worked with other specialists such as behavioral scientists, implementation scientists, and end-users to turn medical data into actionable inferences. He emphasized the need for more humans, particularly medical practitioners when building machine learning models in clinical care. To mention a few benefits, medical practitioners give useful feedback early in the loop and provide critical advice on the features that are most useful for the analysis. Ravi corroborated his position with experiences he had in illness prognosis in the clinic.
Speaking about his article, How the Pandemic Made Algorithms Go Haywire, Ravi talked about experiences that made him realize many machine learning models needed a recheck after the pandemic. He also spoke about how medical practitioners can identify haywire algorithms post-COVID and remedy them. One of the strategies he mentioned involved a quarterly review of the model’s performance. Ravi delved deeper into the activities involved when performing this review.
To put things into perspective, he highlighted data features in cancer screening that were susceptible to changes due to the pandemic. This was a specific example but applies to many other medical applications. He also spoke on ways to make machine learning models for medical use cases more robust to changing times. His methodology involved self-learning models and automatic feature engineering.
Listen to the episode at Data Skeptic.