Diagnosing atypical parkinsonism ante-mortem is—for the most part—a clinical enterprise based mostly on history and the neurologic exam. Diagnostic conundrums are a frequent occurrence given that signs and symptoms overlap not only with other kinds of parkinsonism (i.e., idiopathic; genetic; secondary), but also within the family of atypical parkinsonian syndromes. Establishing a firm—though not yet definitive pathological—diagnosis may take years, imposing a high burden on both patients and health systems [18], particularly in resource-limited countries. Against this background, we deemed it important to evaluate diagnostic pitfalls and develop diagnostic tools for the use of general neurologists.

To the best of our knowledge, the study herein presented is the first original research into atypical parkinsonism diagnostic accuracy in Romania. In addition, to the best of our knowledge, no diagnostic algorithm has hitherto been put together that could help improve the diagnosis of atypical parkinsonism. The results of this study indicate that diagnostic work can be improved at several levels as will be detailed below.

First, while a Boolean-logic tool might be in place, its use is predicated on the full and correct understanding of the particular meaning of signs and symptoms against which a diagnosis is checked. Our initial data-combing through the clinic’s database at the time of patient inclusion revealed that clinicians were not always on the lookout for all signs and symptoms associated with atypical parkinsonism. Nor was one and the same term in the criteria described consistently or comprehensively across clinicians. To give only a few examples, we identified some error-prone areas as follows:

• ‑vertical gaze palsy was either over- or under-interpreted (saccade initiation instead of velocity being considered as a marker of the condition; isolated and barely perceptible limitation of up-gaze being over-interpreted; vertical gaze palsy being considered a PSP-pathognomonic sign without due consideration of “atypical” atypical parkinsonism or other diseases);
• ‑macro square-wave jerks were never identified/mentioned;
• ‑myoclonus and dystonia were under-represented;
• ‑orthostatic hypotension was neither properly defined at 3 min nor fully quantified against systolic and diastolic specific criteria; consideration of other potential and more frequent causes of hypotension (e.g., diabetes, medication, hypovolemic states) was not fully documented;
• ‑alternative/exclusionary causes of urinary dysautonomia (prostatic disease, diabetes, pelvic-floor muscle weakness in elderly women) were not always identified/described;
• ‑inquiries were seldom made about associated erectile dysfunction;
• ‑cognitive deficits were not fully qualified (subdomain-specific limitations were not identified);
• ‑fluctuations in cognition and hallucinations were not fully described against criteria requirements; and
• ‑levodopa trials did not comply with the full criteria specifications for dosage and responsiveness.

Second, the analysis of our data indicated a different disease frequency compared to data reported in the literature. More specifically, physicians in the study identified MSA most frequently, with PSP as the second most frequent diagnosis, followed at a considerable distance by DLB and CBD. In contrast, wider prevalence reports point to DLB as a fairly common disease in the elderly (400 cases per 100,000 persons), with MSA and PSP both having a prevalence of 5 to 10 per 100,000 persons, while the prevalence of CBD is about 1 per 100,000 [19]. The reason for this difference may be related to both the size of the lot under consideration (90 patients) and zonal limitations (one city). Thus, while the two tertiary clinics involved in the study are fairly representative for expertise in movement disorders, their location in the capital city of Bucharest might make it difficult for potential patients across the country to call upon their expert services. One other potential reason is that patients with more prominent dementia and little parkinsonism might first access psychiatry services instead of neurology clinics. Moreover, dementia is most often associated with Alzheimer’s disease with little follow-up in search of other potential etiologies. Last but not least, dementia-related codes as per the electronic record system were not investigated, which may have been a limitation of our methodology, although the combination with parkinsonism or other manifestations as secondary diagnoses should have helped with the identification of less obvious cases. A similar comment could be made with respect to idiopathic Parkinson’s disease, which was not included in the original diagnostic code search. Had this been the case, disease frequencies might have looked different given that atypical parkinsonism is often mistaken for idiopathic Parkinson’s disease [7].

Third, this physician-algorithm comparative study revealed a fairly high rate of correct diagnoses at 56% of the total. However, as many as 23% of all cases were misdiagnosed, which is a rather significant proportion of failure. The remainder (21%) is made up of incomplete diagnoses reflecting perhaps the grey zones of clinical overlap. Diagnostic dilemmas are frequent both among senior doctors and among junior specialists. The reasons for this overall picture are, perhaps, manifold, as follows:

1. consensus criteria are rather lengthy and difficult to follow; they also are not readily available in one place/on-the-go;
2. time/resource limitations during one or several office visits may detract from full consideration of the criteria;
3. symptom-/sign-related terms are not always fully defined and usage is rather specialist-oriented;
4. criteria include several areas of overlap among the four diseases with diagnostic choices difficult to make when symptoms and signs are barely manifest;
5. some consensus criteria (MSA and DLB) do not include absolute exclusion criteria that might have helped rule out potential alternatives at an earlier stage; in fact, the MSA criteria have been fairly recently criticized [20] in light of the significant mismatch between clinical and pathological diagnoses, with a call for criteria revision that is currently under way; as far as DLB is concerned, the debate is not yet over as to a real biological difference between DLB and Parkinson’s disease dementia as opposed to both entities being situated on a disease/pathologic continuum [21]; and
6. physicians do not seem to be applying the mandatory or context-dependent exclusion criteria in place for PSP and CBD.

Fourth, it could be noted from the analysis that there were many nuances within the major diagnostic categories. Among these, the most frequent within the “correct” category seemed to be a failure to fully characterize the disease by including specifications on phenotypes and degrees of certainty. One of the reasons for this failure might be that physicians tend to ignore treatment-irrelevant details and do not think in broader research-related terms as opposed to a more limited, immediate, clinical relevance. However, as has amply been made clear, phenotype and degree of certainty differentiation are critical not only in terms of prognosis but also in terms of clinical trial recruitment [4,22].

To add finesse to the point highlighted above, one might also note that PSP was the one disease with the widest dispersion of diagnostic precision and with most instances of failure to qualify the relevant phenotype. This may appear paradoxical, given that PSP consensus criteria have been fairly recently revised specifically with a view to enhancing diagnostic precision. In our local practice, however, more specific criteria did not translate into better phenotypic identification. One potential reason for this unexpected outcome might be that real-life clinical practice does not allow practitioners to go at length through all the relevant items.

Along the spectrum of tauopathies, CBD criteria are less sophisticated than PSP criteria; however, most CBD diagnoses fell within the “incomplete” category deriving from instances of PSP/CBD overlaps. The reason for this might be that no criteria are available to delineate between these two phenotypically and pathologically similar diseases. In fact, the degree of overlap can be such that some specialists in the field question the clinical utility of a separation between the two [23]. Furthermore, this has led to the proposal that a joint “tauopathy 4R” diagnosis be used in cases of significant overlap [4,24]. CBD is also the least frequent correct diagnosis and misdiagnosis when no criteria for CBD are met. This could be explained by physicians rarely taking CBD into account, except only when a striking CBS phenotype is at stake, coupled with a failure to apply criteria for other CBD phenotypes.

Compared to the wide PSP diagnostic dispersion, MSA seemed to be most often correctly diagnosed. The reason for this might be a combination of criteria simplicity and high frequency of dysautonomia in the elderly, which raises a red flag for the early consideration of an MSA diagnosis. The latter might also be the reason why other atypical parkinsonism cases are often categorized as MSA. Along the alpha-synucleinopathy spectrum, DLB was associated with the highest rate of incomplete and incorrect diagnoses (although this tentative analysis might be skewed by the limited number of DLB cases under consideration—8 cases in total). Still, one explanation might derive from a failure to identify cognitive fluctuations and construe hallucinations as defined in the relevant criteria.

For all these reasons, we have thought it important to make a first attempt at developing an algorithmic diagnostic tool for atypical parkinsonism that could strike the right balance between scientific complexity and simplicity of use for the benefit of regular clinicians. The encouraging results of this retrospective, file-based, study lay the groundwork for future efforts at turning this currently Excel-based tool into a full-fledged smart application that will need proper validation in a properly sampled, multicenter prospective study. It is expected that the code of such an application will be flexible enough to accommodate any forthcoming criteria revisions as well as other relevant criteria for differentials with related groups of diseases. Moreover, we trust the results of this study could help configure a national registry of atypical parkinsonism in the interests of patients, clinicians and researchers alike.

The simple diagnostic approach considered in this article is set against a research background that is rife with attempts at artificial intelligence/machine learning-based diagnostic tools. Indeed, efforts are being made for automated detection of eye movement or gait patterns [25,26]—to give only two examples. For the time being, however, such efforts address the automated recognition of disparate signs and symptoms (saccades, gait, etc.) included in the consensus criteria. Such machine outputs are yet to be perfected and still rely on human trial-and-error input for satisfactory sign/symptom recognition. The merit of our physician-driven diagnostic approach is that it considers all atypical parkinsonism consensus criteria simultaneously, allowing for the best clinical choice among—sometimes overlapping—diagnostic probabilities. These pathologies are too intricate and data-laden for any single machine-based algorithm to capture their underlying complexity and return an accurate, clinical, ante-mortem diagnosis.

While it could be argued that machines might replace physicians when it comes to diagnostic tools in the future, the debate is still on as to whether the replacement will be total, partial or rather symbiotic [27,28].

At the same time, the approach herein aggregated has its limitations, and there are multiple levels where it may be imprecise in regard to diagnostic accuracy.

First, it is important to recognize that the application of these criteria might yield divergent results depending on the stage of the illness. For example, a validation study found that applying the criteria for probable PSP in the first 3 years of disease resulted in a sensitivity of 33.3% and a specificity of 81.6%, while applying those same criteria more than 3 years after onset had a sensitivity of 81.1% and a specificity of 91.4% [17]. Since our approach tries to apply these criteria as best as possible, this primarily means that they also mirror the imprecision inherent in the dynamics of the disease before the relevant clinical picture is fully blown. In particular, the patients in our study had a mean time to diagnosis of 2.8 years, with a median of 2 years—comparable to reports in other studies on the topic [29].

Second, when combing the four sets of criteria for MSA, DLB, PSP and CBD, yet another level of diagnostic imprecision comes up, inherent in the many areas of overlap among particular criteria subsets, coupled with a step-wise consideration of diagnostic certainty. Our “final” diagnosis was the one indicated by the highest degree of certainty (“probable” > “possible” > “suggestive of”). This is because a probable diagnosis increases specificity for that particular disease (e.g., more than 3 years into the disease, “probable” PSP has a specificity of 91.4%, while “suggestive of” PSP has a specificity of 31.4%) [17].

Third, the entire approach is predicated on how criteria are understood and applied by their very users. However, a tool that aggregates all criteria in one place might direct physician efforts to the comprehensive interpretation and consideration of those criteria as compared to their application based solely on human memory. As such, there would be less room for diagnostic error, at least with respect to the application of formal consensus criteria.

Finally, it should also be borne in mind that a “definitive” diagnosis can only be made based on brain biopsies and that, for the time being, the intricacies of neurodegeneration only allow for, at most, a “probable” diagnosis to be made during regular clinical practice.