Validation of a Parkinson Disease Predictive Model in a Population-Based Study

Abstract

Parkinson disease (PD) has a relatively long prodromal period that may permit early identification to reduce diagnostic testing for other conditions when patients are simply presenting with early PD symptoms, as well as to reduce morbidity from fall-related trauma. Earlier identification also could prove critical to the development of neuroprotective therapies. We previously developed a PD predictive model using demographic and Medicare claims data in a population-based case-control study. e area under the receiver-operating characteristic curve (AUC) indicated good performance. We sought to further validate this PD predictive model. In a randomly selected, population-based cohort of 115,492 Medicare beneficiaries aged 66-90 and without PD in 2009, we applied the predictive model to claims data from the prior five years to estimate the probability of future PD diagnosis. During five years of follow-up, we used 2010-2014 Medicare data to determine PD and vital status and then Cox regression to investigate whether PD probability at baseline was associated with time to PD diagnosis. Within a nested case-control sample, we calculated the AUC, sensitivity, and specificity. A total of 2,326 beneficiaries developed PD. Probability of PD was associated with time to PD diagnosis (p < 0.001, hazard ratio � 13.5, 95% confidence interval (CI) 10.6-17.3 for the highest vs. lowest decile of probability). e AUC was 83.3% (95% CI 82.5%-84.1%). At the cut point that balanced sensitivity and specificity, sensitivity was 76.7% and specificity was 76.2%. In an independent sample of additional Medicare beneficiaries, we again applied the model and observed good performance (AUC � 82.2%, 95% CI 81.1%-83.3%). Administrative claims data can facilitate PD identification within Medicare and Medicare-aged samples.

Identification of Prodromal Presentations of Parkinson’s Disease among Primary Care Outpatients in Germany

“…Although the OR decreased as the time before diagnosis increased, they were already increased >5 years before the formal diagnosis of Parkinson's disease. This has also been found consistently in other studies [12,21,22]. While both presentations are included in the core features of Parkinson's disease [23], primary care physicians may not make a formal diagnosis until there is greater clinical certainty [24].…”

A comparison of prediction approaches for identifying prodromal Parkinson disease

“…Identification of people with PD during the prodromal period represents an urgent research priority due to the need to implement neuroprotective therapies earlier in the neurodegenerative process and to prevent disease related morbidity associated with treatable motor symptoms. Our recent, complementary study [ 14 ] validated the previous PD predictive model [ 13 ], providing evidence that the model is effective and a possible strategy to identify those in the prodromal stage of PD. The current study continues to build upon this work by assessing the value of adding medication data from Medicare Part D to an ICD9/procedure code-based predictive model, as well as applying machine learning methods to further validate and enhance our previous work [ 13 , 14 ].…”

“…Our recent, complementary study [ 14 ] validated the previous PD predictive model [ 13 ], providing evidence that the model is effective and a possible strategy to identify those in the prodromal stage of PD. The current study continues to build upon this work by assessing the value of adding medication data from Medicare Part D to an ICD9/procedure code-based predictive model, as well as applying machine learning methods to further validate and enhance our previous work [ 13 , 14 ]. The current study suggests prescription medication data would not improve performance of our original predictions had pharmacy data been available for all of the beneficiaries in that sample, because the AUCs between the models with and without pharmacy data were quite similar and not statistically different.…”

“…The current study builds upon our previous work by considering whether the addition of prescription medication data improves discrimination and whether a random forest classifier could perform better or help improve the original penalized regression approach [ 13 ]. Attempting to improve the model is the logical next step, since we recently validated our original predictive model in a population-based sample followed forward for PD [ 14 ]. We hypothesized that inclusion of prescription medication data would improve model performance for four reasons: 1) these medication data offer an alternative way to capture information available from diagnosis codes, which could be incomplete; 2) medication data might provide diagnostic confirmation and evidence of disease severity; 3) medications might serve as proxies for biologic pathways that might be predictive of PD; and 4) some medications might increase or decrease risk of PD, regardless of the indication for the medication, and thus could be independently predictive.…”

Herpesvirus Infections and Risk of Parkinson’s Disease

“…These are members of other herpesvirus subfamilies that do not specifically target neuronal cells and were uncommon in our dataset. Second, among the controls in our primary dataset, that is, a random sample without PD, we followed the 115,492 who were alive on January 1, 2010, forward for PD or death through to December 31, 2014 [25]. For both HSV and HZ (but not CMV and EBV, which were too rare), we conducted a survival analysis using Cox proportional hazards regression while adjusting for the same variables as in our primary analysis, as of baseline (January 1, 2010).…”