Precision Medicine (PM) and Artificial Intelligence (AI) algorithms – risk of bias in treatment
June 26th, 2018
Precision medicine is more impactful when we build a story through use of a complete and meaningful dataset. Many publications discuss the idea that precision medicine relies on smart technology and this is true. At Diaceutics we believe that smart technology needs to be paired with the right dataset.
For predictive algorithms in precision healthcare to work well, we need to start with what works well for individuals and deep learning algorithms can be used to classify and categorize data. When we combine deep learning with reinforcement learning – long term self-learning through trial and error – we expect the end results to be meaningful or to tell a story, but that may or may not meet expectations.
Volumes of organized data collected systematically over time allow us to build a story so that we can remark that ‘in patients like this, in the recent past, these are the things that have worked well’. Not all data, however, is collected systematically. There are ethical concerns linked to underrepresentation of patient groups that might need to be addressed e.g. some ethnic/socioeconomic/age groups may be less likely to consent to participating in a data sharing project. Therefore, patient consented data sets risk some patient groups being underrepresented. A predictive algorithm used in a computer aided diagnostic tool, developed through use of such potentially not fully representative data sets, might lead to biased algorithms and therefore benefits for one patient group, prospective harm in an underrepresented patient group and a risk of further inequalities in access to healthcare. A broader more global dataset that captures details about demographics would be ideal.
Clinical laboratory data can be regarded as real-world data and is expected to tell a more complete story in terms of describing patient treatment pathways in the population at large. We know that individual laboratories have a unique view of the populations they serve. Content gathered is patient-centric and linked to biomarkers or surrogate biomarkers.
The benefit of clinical laboratory data used to improve the diagnostic process is invaluable and leads to precision or targeted treatments – getting patients on the right drug at the right time.
The value and impact of big data within the diagnostic and treatment journey of patients is an area of interest and a greater focus needs to be placed on how this data can be used.
An AI enabled care pathway can certainly lead to faster and more reliable follow up for some patients, so long as we ensure that the dataset is meaningful in the design of the diagnostic and treatment algorithms and improves healthcare for the many not the few.