Read on Twitter
Predicting the likelihood of future #psychiatricdisorders based on training a #MachineLearning #ML #AI model using a handful of genetic variants is unsound science, and likely to lead to medical discrimination. Thread 1/9. Preprint: https://bit.ly/2GSjiq6
What makes such attempts even more problematic is the risk that people will be denied needed pain relief, and potentially experience stigma associated with a substance use disorder ( #addiction) or other #psychiatricdisorders. 2/9
We demonstrate this examining an attempt to predict future Opioid Dependence (OD). @AlexanderHatoum @PGCgenetics 3/9 https://bit.ly/2GSjiq6
The company’s claim is that #ML classifiers trained with gene variants purportedly from brain reward pathways will predict with very high likelihood which patients are likely to become opioid dependent before prescription, to alert the physician. 4/9
We know from many years of psychiatric genetics that while genetics contributes about half of the risk of opioid dependence. And a small number of variants could not capture much of the genetic risk. 5/9
We demonstrate in the paper why the claims made by one company are not meaningful. tl:dr: the models are confounded by ancestry ( #AlgorithmicBias) and the chosen SNPs perform no better than a random set. https://bit.ly/2GSjiq6 @AlexanderHatoum @PGCgenetics 6/9
The chosen "brain reward system" SNPs vary in MAF among populations, and machine learning models based on them predict ancestry better than case status, when tested with carefully assessed samples. Random SNPs with similar MAF behave the same. https://bit.ly/2GSjiq6 7/9
The results in fact predict African ancestry, not risk for opioid use disorder. They could potentially exacerbate medical discrimination against minority populations. 8/9
Although our study used one example of prediction (opioid use disorder) similar issues are likely for other psychiatric disorders, a cautionary note as people look toward precision genomic medicine in psychiatry. https://bit.ly/2GSjiq6 @AlexanderHatoum @PGCgenetics 9/9
