Refining the diagnostic accuracy of Parkinsonian disorders using metaphenomic annotation of the clinicopathological literature.

The diagnostic precision of Parkinsonian disorders is not accurate enough. Even in expert clinics, up to one in five diagnoses are incorrect. Gold standard diagnosis is post-mortem confirmation of the underlying proteinopathy; however, many clinicopathological studies focus on either a single disease or frame analyses in one temporal direction that may underestimate the true extent of mis- and missed diagnoses. We identified 125 published clinicopathological studies since 1992, extracted phenotype information for ~9200 post-mortem cases, curated the data in a standardised machine-readable format and used this to develop a probabilistic model to quantify diagnostic likelihood based on clinical observations. We found diagnostic accuracy was highest for multiple system atrophy (MSA, 92.8%) and lowest for dementia with Lewy bodies (DLB, 82.1%). MSA and progressive supranuclear palsy were most frequently mis-labelled as Parkinson's disease (PD) in life (7.2% and 8.3% of cases), whereas the most common PD misdiagnosis was Alzheimer's (~7% cases). We calculated likelihood ratios for a large range of clinical phenotypes and demonstrated how these can be used to help refine and improve diagnostic accuracy. This work delivers a harmonised, open-source dataset representing over 30 years of published results and represents a key foundation for flexible predictive models that leverage different sources of information to better discriminate Parkinsonian disorders during the early and prodromal phases of the illness.