An algorithm as a diagnostic tool for central ocular motor disorders, also to diagnose rare disorders

Background Recently an increasing number of digital tools to aid clinical work have been published. This study’s aim was to create an algorithm which can assist physicians as a “digital expert” with the differential diagnosis of central ocular motor disorders, in particular in rare diseases. Results The algorithm’s input consists of a maximum of 60 neurological and oculomotor signs and symptoms. The output is a list of the most probable diagnoses out of 14 alternatives and the most likely topographical anatomical localizations out of eight alternatives. Positive points are given for disease-associated symptoms, negative points for symptoms unlikely to occur with a disease. The accuracy of the algorithm was evaluated using the two diagnoses and two brain zones with the highest scores. In a first step, a dataset of 102 patients (56 males, 48.0 ± 22 yrs) with various central ocular motor disorders and underlying diseases, with a particular focus on rare diseases, was used as the basis for developing the algorithm iteratively. In a second step, the algorithm was validated with a dataset of 104 patients (59 males, 46.0 ± 23 yrs). For 12/14 diseases, the algorithm showed a sensitivity of between 80 and 100% and the specificity of 9/14 diseases was between 82 and 95% (e.g., 100% sensitivity and 75.5% specificity for Niemann Pick type C, and 80% specificity and 91.5% sensitivity for Gaucher’s disease). In terms of a topographic anatomical diagnosis, the sensitivity was between 77 and 100% for 4/8 brain zones, and the specificity of 5/8 zones ranged between 79 and 99%. Conclusion This algorithm using our knowledge of the functional anatomy of the ocular motor system and possible underlying diseases is a useful tool, in particular for the diagnosis of rare diseases associated with typical central ocular motor disorders, which are often overlooked. Electronic supplementary material The online version of this article (10.1186/s13023-019-1164-8) contains supplementary material, which is available to authorized users.


Background
Clinical practice shows that the diagnosis of rare diseases and central ocular motor disorders is often difficult, even for neurologists. On the other hand, we do have detailed knowledge on the anatomy, physiology and pathophysiology of ocular motor disorders, which allows a precise topographic anatomical diagnosis based on bedside examination even without any laboratory examinations [1] (see Table 3 for a short description of the most important parts of the clinical oculomotor examination). This means that, on the basis of clinical information, we can determine whether there is an impairment in the midbrain, pons, medulla or the cerebellar flocculus, nodulus, vermis, or fastigial nucleus.
Rare diseases, such as Niemann-Pick type C (NPC) [2], Tay-Sachs (TS) or Gaucher's disease type 3 (GD 3), are often overlooked, although the diagnosis can often be made on the basis of the patient history and clinical examination and confirmed by genetic testing. Several of these diseases are characterized by quite specific ocular motor findings, such as a supranuclear saccade orat a later stage of the diseasegaze palsy in NPC and TS (for reference see [1]). From a therapeutic point of view, these diseases should also not be overlooked because several of them are treatable nowadays [3,4].
Facing these problems, we designed a simple and easyto-use algorithm to help clinicians to correctly diagnose central ocular motor disorders and, in particular, associated rare diseases. Similar approaches have been recently used to diagnose cerebellar ataxias [5] or vertigo and dizziness [6].

Methods
The algorithm was created in three steps.
Step one Two lists were designed: list A contained 14 diseases which often present with ocular motor disorders, list B contained 60 signs and symptoms typically found in these diseases. The latter can be subdivided into two major groups: general and ocular motor signs and symptoms (see Additional file 1).
Subsequently a table with list A in the cross column and list B in the along column was developed. Based on the current literature [1,7,8], we linked the symptoms to the diseases by simply entering "Yes" if the symptom occurs with the disease and "No" if it does not.
The basic working principle of the algorithm was to create a score for all of the 14 diseases as an output following the input of a patient's signs and symptoms. The symptoms are entered into an entry mask with "Yes" if the patient suffers from a symptom, "No" if he does not and "0" if a symptom was not tested or not testable (see Additional file 3).
The algorithm was further improved by increasing the strength of the linking of very typical symptoms to certain diseases. In the above-mentioned table we entered not "Yes" but "HR" for "highly related". If this symptom occurred, two points instead of one were added to a disease's score. We implemented this linking with, e.g., "internuclear ophthalmoplegia, aged < 60 years" and "MS", "vertical saccade palsy" and "NPC", "resting tremor" and "Parkinsonian syndromes". We also implemented a negative linking meaning that if a certain symptom occurs, the score of a disease was decreased. If, for instance, "paresis" occurs, the score of "NPC" and "GD3" is decreased by two points to better differentiate it from "TS".
Step two The first version of the algorithm was improved using the data from 102 patients (56 males, 48.0 ± 22 yrs., distribution of the diseases: NPC -7, AT -5, AOA1,2-5, GD3-7, TS -5, Wernicke encephalopathy -5, Huntington's chorea -6, MS -10, Parkinson syndromes -9, PSP -9, tumor -4, infarction/hemorrhage -9, inflammatory encephalitis -5, various cerebellar syndromes -16). Most of these patients had been examined at our University Hospital in the past, independently of this study [3,17]. We went through the documented oculomotor examinations and looked for patients who fulfilled our criteria. There were two inclusion criteria: 1. they had to be diagnosed with one and only one of the diseases in list A, and 2. they had to have oculomotor disorders which were found and described exactly in the documentation of the examination. The following exclusion criterion applied: patients had not to have had a second condition causing oculomotor disorder, such as brain surgery or a stroke in the past.
We put the clinical findings from these patients into the entry mask of the algorithm and evaluated its output. Then we adjusted the algorithm in an iterative way until we reached a good sensitivity and specificity. The arithmetic procedures we used in the algorithm were adding zero, one, two, three or four points to the score or subtracting one, two or three points.
We used the same approach as described above to make the algorithm produce a suggestion on the topographical anatomical localization of the lesion. List B with the symptoms remained exactly the same, while list A with the diseases was changed into a list of brain zones, which, when affected, result in ocular motor disorders. Again we used current literature to link the symptoms to the eight zones: midbrain, pons, medulla oblongata, basal ganglia, frontoparietal cortex and the three parts of the cerebellum flocculus/paraflocculus, vermis/fastigial nucleus and nodulus/uvula [8] (see Additional file 2).
We postulated three rules for interpreting the algorithm's result for the diseases: 1. The result consists of the two diseases which get the highest scores in the output list (see Additional file 4). This can be more than two diseases if several get the same score. 2. If the algorithm provides more than five diseases as the result, we considered this as not helpful. When calculating the diseases' sensitivity and specificity we counted such results as false negatives for the actual disease and as false positives for the other 13 diseases. 3. If one disease's score was at least three points higher than any other score, this disease was considered as the only result of the algorithm. When the correct diagnosis appeared in the above-defined result of the algorithm consisting of one to five diseases we counted the result as a true positive for the actual disease and a true negative for the other diseases that did not appear in the result. Every incorrect one of the one to five result-diseases was counted as a false positive.
To interpret the algorithm's result for the topographic anatomical location we also postulated three rules similar to but not identical to the rules for the disease: 1. The result consists of the two brain zones which get the highest scores in the algorithm's output list. This can be more than two zones if several get the same score. 2. Every score with only one point or less is ignored unless one point is the highest existing score. 3. If the algorithm provides more than four zones as a result, we considered this as not helpful and treated it as mentioned above. The sensitivity and specificity were calculated in the same way as for the diseases described above.
Approval from the ethics committee board of the University of Munich was obtained for the study. All investigations were conducted according to the principles of the Declaration of Helsinki.

Statistical analysis
For the statistical evaluation, the software "SAS" v9.3 was used. We calculated the confidence limits of the sensitivity/specificity using an asymptotic normal approximation to the binomial distribution. The whole algorithm was then embedded in an easy-to-use web tool which can be seen in Fig. 1 (called ADOC -Algorithm for the Diagnosis of OCulomotor disorders).

Results
As mentioned in Methods, the algorithm to diagnose the affected brain structures and diseases was developed in an iterative way. In the following, the sensitivity and specificity are given for the last version.
As our result design consists of at least two suggestions about the underlying disease in most cases, there was at least one false positive in every output. So, as expected, the specificity was not as high, ranging from 96 to 63% (best: infarction/hemorrhage and Parkinsonian syndromes 96%, Wernicke's encephalopathy 95%; and MS 63%).

Discussion
The major findings of this study are as follows: First, this algorithm can be a helpful tool for diagnosing, in particular, rare diseases associated with central ocular motor disorders. For example, in the validation cohort we reached a sensitivity of 100% for NPC (10/10) and Wernicke's encephalopathy (5/5). It is Fig. 1 Screenshot of the data entry mask in the finished web tool. This excerpt shows the main signs and symptoms categories of the data entry file. By clicking on "Yes" or "No" one confirms or denies a symptom. Symptoms that were not tested can just be skipped by not clicking on any of the possibilities and leaving the field empty Table 1 Sensitivity and specificity for the diseases in the validation cohort. Sensitivity ranged from 100% for NPC, AOA1 and 2, TS, Wernicke's encephalopathy, inflammatory encephalitis, infarction/hemorrhage to 60% for AT. Specificity was between 95% for Parkinsonian syndromes and Huntington's chorea and 66% for inflammatory encephalitis. Additionally the 95% confidence interval was calculated for every value  Table 2 Sensitivity and specificity for the brain zones in the validation cohort. Sensitivity ranged from 100% for medulla oblongata to 0% for nodulus/uvula. Specificity was between 99% for frontoparietal cortex and 52% for midbrain. Additionally the 95% confidence interval was calculated for every value  assumed that both of them are vastly underdiagnosed [11,19]. Since these diseases are treatable or, in the case of Wernicke's encephalopathy, even curable, an early diagnosis has a huge impact on the outcome of these patients. Second, the results for the brain zones were generally worse but can still give an indication of where to look for pathologies in imaging. In the validation cohort, the sensitivity for involvement of the medulla oblongata was 100% (4/4) and for the pons 82.4% (28/34). Third, the algorithm can be applied in less than 5 min.
Compared to "medx" [6], a similar tool recently published to diagnose vertigo and dizziness, our algorithm showed a higher sensitivity (medx: 40 to 80.5%) but a lower specificity (medx: at least 80%). This can perhaps be explained by the fact that "medx" focuses on the first suggested diagnosis, whereas our tool presents the two top-scoring results. Since our algorithm deals with more rare diseases, the different approaches seem to be suitable for the different problems they are supposed to solve. Another recent algorithm to diagnose recessive ataxias is called "RADIAL" [5]. It showed a higher average sensitivity and specificity (RADIAL: 92.2 and 95.4%, respectively) than our tool but it works with around twice as many features (120 vs. 60).
This study has several limitations: First, it was a retrospective analysis. Second, our gold standard was the diagnosis made in the hospital, which is not flawless. Third, a major problem was that the affected brain zones could not always be verified in the brain imaging available or that patients had multiple lesions as in MS. Regarding the cerebellum, imaging often shows no pathologies, however the clinical signs are often specific based on current knowledge of the function and dysfunction of the flocculus/paraflocculus, nodulus, nucleus fastigii and dorsal vermis. All in all, however, the major focus was on the diagnosis of rare diseases which can evidently be improved by such a simple algorithm.

Conclusions
In summary, this algorithm uses our knowledge on the functional anatomy of the ocular motor system. It is based on the simple idea of comparing signs and symptoms typical of certain diseases and brain lesions to signs and symptoms occurring in a certain patient. It is a useful tool for diagnosing diseases, in particular rare ones, which present with central ocular motor disorders.