Improving Diagnosis of Rare Genetic Diseases Through Matching Markets
http://med.stanford.edu/news/all-news/2018/08/new-algorithm-could-improve-diagnosis-of-rare-diseases.html
With the genetic information available to us today, detecting genetic diseases is easy in theory. But matching the disease to the symptoms and the genetic code isn’t so easy for doctors everyday, especially when you’re looking at rare diseases. This article discusses a new algorithm, called “Phrank” (phenotype+rank) created by Gill Bejerano’s team at Stanford University, that matches the patient’s symptoms and genetic information to most likely diagnoses by sorting through a database of medical literature. While a clinician could spend 20-40 hours to arrive at a diagnosis, Phrank takes 1-4, and uses no human labor. Phrank consistently has performed well; it ranks the correct diagnosis fourth in its list on average, as compared to previous algorithms that performed significantly worse. Phrank also can help doctors discover new genetic diseases, when symptoms do not match to literature. Bejerano, the head researcher, argues that “nobody is going to replace a clinician making a diagnosis”, but that this tool makes it significantly easier for physicians.
This ties into what we’ve discussed in class in that this sort of matching parallels what happens in matching markets. The people are the patients, and the goods are the potential diagnoses. While a perfect matching in a matching market means that each patient gets a different diagnosis, in this case, the ideal case is that each patient gets a correct diagnosis. Since there is no perfect matching, there must be a constricted set, based on the Matching Theorem; since there are multiple possible diagnoses for each patient, a constricted set exists. The idea of valuations in matching markets also applies here; each diagnosis is assigned a different probabilistic value based on how well it matches the genetic information and symptoms. The higher the probability of each diagnosis, the more likely it is the correct one, but, as discussed above, the algorithm is not perfect, so this might not be the case. Therefore, the relationship between correctness of the diagnosis (the true value of the algorithm) and the probability assigned by the algorithm (what is being used as valuation) is not straightforward, but the algorithm must rely on the idea that the higher the probability score, the better match the diagnosis makes to the patient. It is interesting to contemplate the similarities and differences between this sort of matching and the matching markets we discussed in class. It certainly does maximize social welfare when patients get diagnosed correctly and efficiently!