Translational Vision Science & Technology Cover Image for Volume 14, Issue 6
June 2025
Volume 14, Issue 6
Open Access
Artificial Intelligence  |   June 2025
Machine Learning Applied to Visual Fields of Dominant Optic Atrophy Patients
Author Affiliations & Notes
  • Catarina P. Coutinho
    Department of Pharmacy and Biotechnology, University of Bologna, Bologna, Italy
    Studio Oculistico d'Azeglio, Bologna, Italy
  • Ferdinando Zanchetta
    Department of Pharmacy and Biotechnology, University of Bologna, Bologna, Italy
  • Michele Carbonelli
    Department of Biomedical and Neuromotor Sciences, University of Bologna, Bologna, Italy
  • Marco Battista
    Department of Ophthalmology, University Vita-Salute, IRCCS Ospedale San Raffaele, Milan, Italy
  • Alice Galzignato
    Studio Oculistico d'Azeglio, Bologna, Italy
  • Chiara La Morgia
    Department of Biomedical and Neuromotor Sciences, University of Bologna, Bologna, Italy
    IRCCS, Istituto delle Scienze Neurologiche, Programma di Neurogenetica, Bologna, Italy
  • Giulia Amore
    Department of Biomedical and Neuromotor Sciences, University of Bologna, Bologna, Italy
  • Martina Romagnoli
    IRCCS, Istituto delle Scienze Neurologiche, Programma di Neurogenetica, Bologna, Italy
  • Giacomo Savini
    G. B. Bietti Foundation IRCCS, Rome, Italy
  • Luigi Brotto
    Department of Clinical Science and Community Health, University of Milan, Milan, Italy
  • Paolo Nucci
    Department of Clinical Science and Community Health, University of Milan, Milan, Italy
  • Leonardo Caporali
    IRCCS, Istituto delle Scienze Neurologiche, Programma di Neurogenetica, Bologna, Italy
  • Francesco Bandello
    Department of Ophthalmology, University Vita-Salute, IRCCS Ospedale San Raffaele, Milan, Italy
  • Valerio Carelli
    Department of Biomedical and Neuromotor Sciences, University of Bologna, Bologna, Italy
    IRCCS, Istituto delle Scienze Neurologiche, Programma di Neurogenetica, Bologna, Italy
  • Maria Lucia Cascavilla
    Department of Ophthalmology, University Vita-Salute, IRCCS Ospedale San Raffaele, Milan, Italy
  • Rita Fioresi
    Department of Pharmacy and Biotechnology, University of Bologna, Bologna, Italy
  • Piero Barboni
    Studio Oculistico d'Azeglio, Bologna, Italy
    Department of Ophthalmology, University Vita-Salute, IRCCS Ospedale San Raffaele, Milan, Italy
  • Correspondence: Piero Barboni, IRCCS San Raffaele Scientific Institute, Via Olgettina, 60, Milan 20132, Italy. e-mail: [email protected] 
  • Footnotes
     CPC, FZ, RF, and PB contributed equally to this work.
Translational Vision Science & Technology June 2025, Vol.14, 20. doi:https://doi.org/10.1167/tvst.14.6.20
  • Views
  • PDF
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Catarina P. Coutinho, Ferdinando Zanchetta, Michele Carbonelli, Marco Battista, Alice Galzignato, Chiara La Morgia, Giulia Amore, Martina Romagnoli, Giacomo Savini, Luigi Brotto, Paolo Nucci, Leonardo Caporali, Francesco Bandello, Valerio Carelli, Maria Lucia Cascavilla, Rita Fioresi, Piero Barboni; Machine Learning Applied to Visual Fields of Dominant Optic Atrophy Patients. Trans. Vis. Sci. Tech. 2025;14(6):20. https://doi.org/10.1167/tvst.14.6.20.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Purpose: Identification and quantification of characteristic visual field (VF) patterns in patients with dominant optic atrophy (DOA) using the archetypal analysis (AA) machine learning algorithm.

Methods: In this retrospective study, we collected 30-2 or 24-2 VFs performed with Humphrey Visual Field analyzer from 144 patients (280 eyes) affected by molecularly confirmed DOA carrying OPA1 heterozygous mutation. The VFs were randomly separated into a training set (224 VFs, 80%) and test set (56 VFs, 20%). An AA model was developed by decomposing the VFs of the training set into archetypes (ATs). Spearman correlations were calculated between ATs’ weights and mean deviation (MD) and visual acuity (VA). Statistical comparison was performed between ATs weights according to mutation subtype groups.

Results: The DOA-AA model was composed of eight ATs with a high performance in the test set (R2 = 0.88). According to the Ocular Hypertension Treatment Study (OHTS) classification, the central/ceco-central scotoma resembling ATs presented the highest weights (24%) followed by superior defects (13%). ATs with more abnormal VF resembling defects correlated most with MD (AT5-8), whereas only the total loss AT7 with VA (P value < 0.01). Subtype mutations linked with worse clinical features had statistically significantly higher weights for worse ATs (AT7, P < 0.001).

Conclusions: The developed AA model allowed the identification and quantification of VF patterns in DOA. Furthermore, a clinical genotype-phenotype association was supported by the comparison of severity at VF AA decomposition.

Translational Relevance: AA enables an objective identification of quantifiable visual field defects intrinsic to DOA providing functional details based on genotype.

Introduction
Dominant optic atrophy (DOA) is an inherited disease caused by mitochondrial dysfunction that leads to selective degeneration of retinal ganglion cells (RGCs) and their axons constituting the optic nerve. DOA is a relatively common inherited optic atrophy, with a prevalence of 1 in 35,000 worldwide.13 Most patients with DOA carry a heterozygous mutation in the OPA1 gene, which encodes for a mitochondrial inner membrane protein with multifunctional properties.2,3 Over 400 OPA1 variants deemed pathogenic have been reported and lead to RGCs degeneration through various mechanisms including impaired fusion of the inner mitochondrial membrane implying defective oxidative phosphorylation, increase of reactive oxygen species, and altered calcium homeostasis.46 According to the OPA1 mutation type, we can distinguish those leading to haploinsufficiency with reduced amount of normal protein as due to deletions or splice site/stop mutations, for example, from those being missense variants which are assumed to induce a dominant negative effect. Specifically, the missense OPA1 mutations located within the GTPase domain have been associated with worse clinical features, multisystemic involvement qualifying the DOA “plus” spectrum.1,7,8 
The progressive dysfunction and degeneration of RGCs lead to the characteristic loss of central vision and optic nerve atrophy. The disease is clinically characterized by slowly progressive bilateral visual loss associated with dyschromatopsia, central scotoma, and optic disc pallor. The onset of symptoms is usually in early childhood and the progression of visual loss is typically slow or has a stepwise progression with intercurrent stability.911 The disease is known for its marked variability regarding the age of presentation and clinical expression, particularly in terms of visual acuity (VA) and visual field (VF) defects among affected individuals. This means that the severity and rate of vision loss can vary significantly from person to person, even within the same family with the same genetic mutation. VA can vary from 0.0 LogMAR to light perception, although most often patients present moderate loss of VA in the range of 0.2 to 0.6 LogMAR.12,13 The typical VF pattern is a central scotoma due to the preferential involvement of the papillomacular bundle; however, it can vary widely from subtle defects to subtotal VF loss or include specific patterns such as superotemporal VF loss.4,1014 Nevertheless, no clinical studies so far provided a detailed description of VF and its progression in DOA. Another hallmark of the disease is that functional changes may not correlate directly with the structural changes leading to a poor clinical function to structure correspondence.15 The thickness of macular ganglion cells (ganglion cell with inner plexiform layer [GC-IPL]) measured by optical coherence tomography (OCT) shows a diffuse reduction often more severe than the actual functional impairment, configuring a structural and functional discrepancy.16,17 
To analyze VF loss in an automatized way, different methods have been explored among which the archetypal analysis (AA). This unsupervised machine learning (ML) algorithm has the following important features: (1) it allows the identification of a restricted number of important spatial patterns of VF loss in a given dataset, called archetypes (ATs), (2) it allows the quantification of their proportions in each sample of the dataset or in out of sample data, and (3) it provides the clinicians with interpretable visual representations of the information obtained in (1) and (2). It is important to emphasize that both the resulting ATs and their quantification in each sample of a given dataset do not rely on the consensus of a pool of experts, as the identification of the ATs is performed entirely by the algorithm in an unsupervised manner. In recent studies, AA has already been explored as a mathematical approach for the identification and quantification of VF defects within a single disease such as glaucoma, idiopathic intracranial hypertension (IIH), and optic neuritis (ON).1821 
Furthermore, specific prognostic factors related to visual function beyond VA are lacking. Therefore, identifying patterns of VF loss could help to enhance the recognition of visual impairment in DOA. In this study, we aimed to implement AA for the VF analysis of patients with DOA hypothesizing that the VF ATs obtained would resemble the clinical characteristic patterns of VF loss in a hereditary optic neuropathy. Moreover, we also analyzed the correlation of ATs subgrouping of patients with DOA with the different classes of OPA1 mutations. 
Methods
In this retrospective study, we included patients affected by molecularly confirmed DOA carrying OPA1 heterozygous mutation. The clinical records of eligible patients were reviewed retrospectively for the extraction of clinical and VF data. Patients were evaluated at the San Raffaele Scientific Institute, Milan, and the IRCCS Istituto delle Scienze Neurologiche di Bologna, Bologna, Italy. The study is adherent to the Declaration of Helsinki and was approved by the San Raffaele Hospital, Milan (143/INT/2020) and Bellaria Hospital, Bologna (121/2019/OSS/AUSLBO - 19012), Ethic Committee. Informed consent was obtained from all participants. Ophthalmological phenotyping included assessment of VA, slit-lamp biomicroscopy, intraocular pressure measurement, and indirect ophthalmoscopy. The VA was assessed by measuring the best-corrected visual acuity (BCVA) in Snellen decimals. 
Exclusion Criteria
The exclusion criteria consisted of the presence of any retinal or optic nerve disease other than DOA, patients with spherical or cylindrical refractive errors higher than 5 and 2 diopters (D), respectively; and reliability VF criteria outside normal limit (fixation loss >33%, false-negative rate >20%, and false-positive rate >20%).18,19 
Visual Fields
Multiple VFs performed by SITA standard 30-2 or 24-2 Humphrey VF analyzer (HFA II 750-4.1 2005; Carl Zeiss Meditec, Dublin, CA, USA) were exported in DICOM format. The first VF date with a reliable examination was selected for each case. 
The DICOM files were then imported into Python running in an Anaconda Environment and resorting to the open-source software script hvf_extraction_script,22 VF metadata, value plots, and percentile plots were extracted, the left eyes were converted to the right eye format, and finally all data were exported into an Excel file. A separate Excel file was created only with the total deviations (TDs) where the VFs in a 30-2 format were transformed into a 24-2 format by eliminating 22 TD points (Fig. 1). The TDs corresponding to the blind spot were excluded (TD47 and TD57; see Fig. 1). 
Figure 1.
 
Total deviation plot transformation of a visual field in format 30-2 (74 TD values, TD47 and TD57 represent the blind spot) to a 24-2 format (52 TD values) by excluding the 22 total deviation values represented in red.
Figure 1.
 
Total deviation plot transformation of a visual field in format 30-2 (74 TD values, TD47 and TD57 represent the blind spot) to a 24-2 format (52 TD values) by excluding the 22 total deviation values represented in red.
Archetypal Analysis Implementation
We implemented the Archetypal Analysis package in Python, which is based on the AA model described by Cutler and Breiman.23,24 The TDs were used as data input without being normalized because these represent the deviation to a normal value of an age-matched person.1821 
The dataset was randomly split into a training set (80%) and a test set (20%). The training set was used to determine the optimal number of archetypes via cross-validation and to develop the AA model. The test set did not contain any VFs belonging to the training set. 
To choose the optimal number of ATs, a 10-fold cross-validation was performed, meaning that the shuffled training dataset was partitioned in 10 parts and each subset was used once as the test subset, whereas the remaining 9 subsets were used as the training set. For each fold in the cross-validation, the respective R2 scores were calculated on the test set following Equation 1, where RSS stands for residual sum of squares and is the difference between the actual values and the predicted values, and TSS stands for total sum of squares and is the difference between the actual values and the mean value. In our case, R2 scores can be thought as computing the explained variance.25 The average of the R2 scores were plotted against the number of ATs and the optimal number of ATs was chosen following the Elbow criterion and looking at the R2 scores derivative (Supplementary Material S1).26,27 
\begin{eqnarray}{{R}^2} = 1 - \frac{{RSS}}{{TSS}}\end{eqnarray}
(1)
Once the number of ATs was chosen, the final ATs were computed by fitting an appropriate AA model on the whole training set. All the AT analyses conducted were calculated with 100 random initializations. For each AT analysis conducted, we selected the model corresponding to the best RSS. 
The relative weight (RW) of each AT was computed according to Equation 2 so that the total sum would be 1 (100%) – for each AT i the sum of its RW for all VF TDs j is divided by the sum of all RW (Equation 2).  
\begin{eqnarray}R{{W}_i} = \frac{{\mathop \sum \nolimits_j {{\alpha }_{ij}}}}{{\mathop \sum \nolimits_{ij} {{\alpha }_{ij}}}}\end{eqnarray}
(2)
Here, for a fixed AT i and a fixed VF j, we have denoted αij as the coefficient of the archetype i appearing in the approximated decomposition of the VF j as convex linear combination of the ATs resulting from the developed AA model. 
Visual Fields as a Sum of AT
For the test set, we computed and analyzed the decompositions of different VFs resulting from our AA model by using the ATs of the developed model. Thereby, one VF was described as the weighted sum (convex linear combination) of several ATs within the developed model, whereas higher percentage weights are attributed to those AT patterns that contribute more to the VF reconstruction. 
Statistical Analysis and Composite Score by Doshi et al.20
Considering only the ATs with RW higher than 10% when decomposing the VFs of the training set, Spearman correlations were performed between the ATs RW and the VF's MD and the patient's VA. 
Furthermore, the Composite AT Sum Score was calculated as described by Doshi et al., where normal or less-abnormal VF resembling ATs are attributed a positive sign whereas a negative sign to more abnormal ATs.20 Thereby, the ATs' sum for a single VF reflects all defects rather than just the central ones. Spearman correlations between the Composite AT Sum Score and the VF's MD and the patient's VA were calculated. 
The Kruskal-Wallis test with Dunn's post hoc test were performed for the comparison of ATs’ RW between groups. A P value lower than 0.05 (significance level of 5%) was considered as a statistically significant correlation. 
Mutation Subtype Analysis
To explore if a different VF loss pattern was related to different subtypes of OPA1 mutations, we stratified three groups according to the different mutations: (1) the haploinsufficiency group, (2) the missense non-GTPase group, and (3) the missense GTPase group. For each group, the VFs decomposition was computed using the developed AA model and the RW of each AT was statistically compared between groups. Spearman correlations between the Composite AT Sum Score calculated for each group and the VF's MD and patient's VA were obtained. 
Visual Field Abnormalities Classification Applied to the Archetypes
To further analyze and classify the ATs of the developed DOA AA model, we used two distinct VF abnormalities classification systems. The first one, the Visual Field Consensus, was released in 2022 aiming to uniformize the VF classification by presenting a terminology for the interpretation VF abnormalities.28 Six abnormality groups were defined (neurologic, nerve fiber bundle, central, diffuse, artifactual/retinal, and nonspecific) to classify 24 VF defects. The second system, termed the Ocular Hypertension Treatment Study (OHTS) classification system, focused on the distinction between two groups: never fiber bundle related abnormalities or non-nerve fiber bundle abnormalities, and assigned 17 VF defects into these two groups.29 
Results
We considered 280 eyes of 144 patients, after exclusion of 8 eyes for either not executable or not reliable VF and 21 patients with high myopia (≥ 5 D). Ultimately, the training set was composed of 224 VFs and the remaining 56 VFs were used for testing (Table 1). There was no overlap of VFs between the training and the test datasets. 
Table 1.
 
Demographic Characteristics of Study Cohort
Table 1.
 
Demographic Characteristics of Study Cohort
DOA AA Model
In Figure 2A the average R2 scores versus the number of ATs are plotted for the 10-fold cross-validation. Following the Elbow criterion, we considered eight as the optimal number of ATs because a marked increase is depicted until five ATs, however, only after eight archetypes, the R2 scores grow with a steady linear behavior that we considered to be the terminal one. Figure 2B with the R2 scores difference versus the number of ATs sustained the choice of 8 ATs. 
Figure 2.
 
(A) R2 scores change with the number of archetypes; (B) R2 scores difference against with the number of archetypes.
Figure 2.
 
(A) R2 scores change with the number of archetypes; (B) R2 scores difference against with the number of archetypes.
After the number of ATs was determined to be eight, we ran the AA model on the whole training set (224 VFs) to determine the final eight ATs. The obtained ATs for the DOA AA model are shown in Figure 3, ranked according to their RW in percentages. 
Figure 3.
 
The archetypes obtained with the respective relative weights in percentage. Colors are associated with the total deviations: blue = above normal; and red = below normal (visual field impairment). Points outside the VF are given a dark blue color.
Figure 3.
 
The archetypes obtained with the respective relative weights in percentage. Colors are associated with the total deviations: blue = above normal; and red = below normal (visual field impairment). Points outside the VF are given a dark blue color.
The #AT1 presented the highest weight (48.49%) and represents a normal VF. The #AT2 (12.97%) and #AT3 (10.58%) ATs resemble the most expected DOA defect, a central/ceco-central scotoma. With lower weights, #AT4 and #AT6 present different superior visual impairment patterns: more in detail #AT4 supero-temporal and #AT6 altitudinal scotoma. Widespread visual loss appears in ATs #AT7 and #AT8
Following the abnormalities classification terminology presented in the Visual Field Consensus, the eight ATs of our AA model can be classified as: 
  • Neurologic abnormalities: #AT4 (quadrantanopia) and #AT5 (hemianopia).
  • Nerve fiber bundle abnormalities: #AT6 (altitudinal).
  • Central abnormalities: #AT2 and #AT3 (central/ceco-central).
  • Diffuse abnormalities: #AT7 (total loss) and #AT8 (widespread).
According to this classification, the neurologic abnormalities represent a total of 13.10%, the nerve fiber bundle related abnormalities represent a total of 5.28%, the central ones with the highest total percentage of 23.28%, and, last, the diffuse abnormalities with 9.57%. 
Following the OHTS classification system, six ATs appear to match non-nerve fiber bundle abnormalities (#AT2, #AT3, #AT4, #AT5, #AT7, and #AT8) and only one AT appears to match the nerve fiber bundle defects (#AT6). 
DOA AA Model Implementation and Testing
To evaluate the quality of the ATs we found, we computed the R2 scores obtained after approximating both the training set and the test set. The score on the test set (R2 = 0.88) remained high compared to the score of the training set (R2 = 0.89) and 10-fold cross-validation (R2 = 0.98). 
The 8-AT model was used for the decomposition of the 56 VFs of the test dataset. In Table 2 are presented the RWs in percentages for each AT which represent how much each AT was used to reconstruct the VFs of the test set in comparison to the training set. In addition, in the test set, the most frequent AT was #AT1, followed by the #AT3 that resembles one of the most common DOA patterns, the central scotoma. 
Table 2.
 
Archetype's Relative Weight in Percentage in the Developed Model (Training Set) and for the Decomposition of the VFs of the Test Set
Table 2.
 
Archetype's Relative Weight in Percentage in the Developed Model (Training Set) and for the Decomposition of the VFs of the Test Set
Spearman Correlation and Composite Score by Doshi et al.20
Table 3 summarizes the Spearman coefficients for the correlations between the RWs of each AT and the VF's MD and patient's VA. A statistically significant positive correlation was found between #AT1 and MD whereas significant negative correlations for #AT5, #AT6, #AT7, and #AT8. A strong negative correlation with the VA was found for #AT1 whereas a positive correlation with #AT7
Table 3.
 
Spearman Correlation Coefficients Between the Archetype's Relative Weight and the Mean Deviation, in dB, and Visual Acuity, in logMAR
Table 3.
 
Spearman Correlation Coefficients Between the Archetype's Relative Weight and the Mean Deviation, in dB, and Visual Acuity, in logMAR
A signficant decrease of the Composite AT Sum Score was detected with both MD (Spearman coefficient = 0.8038, P value < 0.001) and VA (Spearman coefficient = −0.3046, P value < 0.001), indicating that the more abnormal ATs (negative sign in the score) tend to correlate with a worse visual performance (Fig. 4). 
Figure 4.
 
Spearman correlation between the Composite AT Sum Score and the mean deviation (dB) on the left side and the visual acuity (logMAR) on the right side.
Figure 4.
 
Spearman correlation between the Composite AT Sum Score and the mean deviation (dB) on the left side and the visual acuity (logMAR) on the right side.
Mutation Subtype Analysis
Considering the 270 cases for which the subtype of mutation classification was available, no statistical differences were found between groups for age (P value = 0.684). A worse MD was found for the missense GTPase group compared to the missense non-GTPase group (P value < 0.001) and a slightly worse MD compared to the haploinsufficiency group (P value = 0.060). VA was better in the missense non-GTPase group compared to the haploinsufficiency group (P value = 0.015) and the missense GTPase group (P value = 0.020). All values are presented in Supplementary Table S2
The statistical comparison of the RW of each AT between the haploinsufficiency, missense non-GTPase, and missense GTPase subtype mutations is presented in Table 4. Although the normal VF resembling AT (#AT1) was statistically higher for the missense non-GTPase group (P value = 0.001), a significantly higher RW of #AT7 (total loss) was found for the missense GTPase group (12.37%) compared with both the haploinsufficiency (4.71%) and the missense non-GTPase (3.22%) groups. #AT3 (central/ceco-central scotoma) had a lower RW in the missense non-GTPase group (5.84%) compared to the other two groups (12.20% and 14.25% for haploinsufficiency and missense GTPase, respectively). In turn, #AT2 appeared to be less frequent in the missense GTPase group (4.99%) compared to the haploinsufficiency group (14.75%), even if not statistically significant. 
Table 4.
 
Archetype's Relative Weight in Percentage for the Haploinsufficiency, Missense Non-GTPase and Missense GTPase Visual Field Decomposition and Global and Post Hoc Test P Values for the Relative Weight Comparison
Table 4.
 
Archetype's Relative Weight in Percentage for the Haploinsufficiency, Missense Non-GTPase and Missense GTPase Visual Field Decomposition and Global and Post Hoc Test P Values for the Relative Weight Comparison
The Spearman correlation revealed a slightly more pronounced statistically significant decrease of the Composite AT Sum Score with the MD for the missense GTPase group (Spearman correlation = 0.9315, P value < 0.001) than for the haploinsufficiency group (Spearman correlation = 0.8161, P value < 0.001) and the missense non-GTPase group (Spearman correlation = 0.7885, P value < 0.001). The same was found for the VA (Haploinsufficiency: Spearman correlation = −0.2517, P value = 0.001; Missense non-GTPase: Spearman correlation: −0.3520, P value = 0.004; and Missense GTPase: Spearman correlation = −0.5113, P value < 0.001; Supplementary Figs. A, B, and C in S2 of the Supplementary Material). 
An example of implementation of the 8-AT model for the left eye of a haploinsufficiency subtype mutation DOA case, with a hemianopia defect, is presented in Figure 5A (−18.08 MD and VA of 1.00 logMAR). The TDs of a 24-2 VF test were used as input and the conversion of the left eye into the right eye format was performed. According to the developed AA model, the patient's VF is decomposed as a convex linear combination of the ATs whose coefficients can be thought of the “proportions” of each AT in it. Thus, using our developed model, this VF can be decomposed as the weighted sum:  
\begin{eqnarray} && {\rm{VF\ }} \approx {\rm{\ }}0.7321{\rm{\ }}\left( {\# AT5} \right){\rm{\ }} \nonumber \\ && + {\rm{\ }}0.2219{\rm{\ }}\left( {\# AT6} \right){\rm{\ }} + {\rm{\ }}0.0461{\rm{\ }}\left( {\# AT5} \right) \end{eqnarray}
(3)
 
Figure 5.
 
Application of the 8-ATs model for the decomposition of: (A) a left eye 24-2 VF and (B) a 30-2 VF regarding the right eye. On the top left, the VF and on the bottom left the respective TDs. On the right are shown the used ATs and their respective RWs for the VF decomposition according to the developed model.
Figure 5.
 
Application of the 8-ATs model for the decomposition of: (A) a left eye 24-2 VF and (B) a 30-2 VF regarding the right eye. On the top left, the VF and on the bottom left the respective TDs. On the right are shown the used ATs and their respective RWs for the VF decomposition according to the developed model.
In Figure 5B, we presented the decomposition of a 30-2 VF regarding the right eye of a patient with missense GTPase subtype mutation DOA with a widespread defect (−28.52 MD and VA of 0.70 logMAR). For this case, the transformation of 30-2 to 24-2 was performed according to Figure 1. The decomposition for this VF according to the developed model is described in Equation 3:  
\begin{eqnarray}{\rm{VF\ }} \approx \ {\rm{\ }}0.8048{\rm{\ }}\left( {\# {{AT}}7} \right){\rm{\ }} + {\rm{\ }}0.1857{\rm{\ }}\left( {\# {{AT}}8} \right)\end{eqnarray}
(4)
 
Discussion
In the present study, we developed an AA model composed of eight ATs that enabled the identification of quantifiable VF loss patterns in DOA. To our knowledge, this is the first work applying AA to a hereditary optic neuropathy. The ATs also allowed to highlight a genotype-phenotype correlation as missense GTPase mutation subtype was significantly distinguishable by the more severe VF and VA impairment. 
In previous studies, this unsupervised ML algorithm provided VF loss ATs comparable to the VF defects described by experts, namely in the OHTS for glaucoma and by experts within Treatment Trials for IIH and ON. In all these cases, AA was able to identify ATs having a clear and interpretable clinical significance and that could in principle be used by other practitioners to run analyses based on them. Indeed, the identification of a candidate set of ATs for a given type of data has an intrinsic significance as they can be used as a common basis for quantitative analyses run by different groups of researchers. In addition, AA allows the decomposition of a single VF into separate patterns of regional VF loss that might be difficult to distinguish with conventional visual field machines. Furthermore, the percentage weight attributed to each AT enables a quantification of the representative VF defect. 
There are very few studies describing VF damage in DOA from the clinical point of view. The most common VF pattern has been typically described as central or ceco-central scotoma. However, Votruba et al. found a superotemporal defect in 66% of cases. Other described VF patterns include enlarged blind spot, bitemporal hemianopia, and subtotal VF loss.11,14 
Following the OHTS VF defect classification, in the presently developed model of eight ATs, the central/ceco-central scotoma had the highest weight (#AT2 and #AT3) followed by the superior defects (#AT4 and #AT6). When we classify the eight ATs according to the VF consensus, the central abnormalities (2 ATs) and the neurological abnormalities (2 ATs) were the most common (23.28% and 13.10%, respectively), thus clinically sustaining the specificity of this DOA AA model; moreover, following the OHTS classification system the non-nerve fiber bundle abnormalities (6 ATs, 46.23%) were the majority, aligning with the previous classification. Hence, these functional results are in agreement with the specific structural damage of small fiber involvement in DOA. In addition, the predominant superior defect points out the DOA typical preferential inferotemporal involvement of RNFL damage.16 
In all models, the normal resembling VF AT has the highest weight due to the decomposition nature of the AA method. However, in the DOA AA model it is even more frequent (48.49%) compared with the others (ON 40.17%, IIH 30.89%, and glaucoma 41.%) which might have different explanations: first, due to the lower number of ATs; second, the DOA VF loss patterns are typically less peripheric than in the other diseases, and, in addition, the central abnormalities have higher structural/functional discrepancy with less VF representation (i.e. a patient with GCL damage might have a normal VF).15 
The significant positive correlation of the normal resembling ATs (#1) with MD along with the negative correlations of the worse ATs (#5–8) with MD were expected because MD is an average of the total deviation from a standard dataset, being therefore insensitive to relatively small scotomas. Instead, the correlations between ATs and VA highlight the different involvement of the central VF: when no scotoma is present also the VA is good (correlation with #AT1), whereas the other extreme when a total scotoma is present also the VA is very low due to the impossibility of eccentric fixation (correlation with #AT7); finally, the lacking of correlation in the intermediate cases (#AT4#AT6, and #AT8) can be explained by the different extensions of central visual field damage that are associated with a sparing peripheral area in which patients present a low/medium vision due to eccentric fixation.12 When considering the AT Sum Scores the correlations become evident: ATs resembling worse VF abnormalities are more present with worse MD and VA. 
The mutation subtype analysis showed a significant association with worse ATs (#7) in the missense GTPase mutation group, whereas better ATs (#1, #2) in missense non-GTPase and haploinsufficiency mutation groups. These findings are in line with previous reports which indicate a worse clinical outcome of missense GTPase mutations.7,8 Further studies will investigate the clinical outcome in non-GTPase mutations which seems to be better. 
In the first example of a haploinsufficiency mutation (see Fig. 5A), the decomposition pattern revealed a major weight of hemianopia AT and less of the others, in contrast, in the second example of a missense GTPase mutation (see Fig. 5B), the major contribution was of the total loss AT with a contribution of central/ceco-central scotoma resembling ATs. This highlights the strength of the AA model to provide detailed VF defect analysis for individual cases. 
The present work focuses on DOA, which being a rare hereditary optic neuropathy has a lower prevalence than other diseases as glaucoma. Therefore, the number of cases enrolled for the AA application was the major limitation in this study. Even so, our model of eight ATs presented high scores both in the training and testing phases, highlighting the potential as a tool for VF loss analysis in DOA. 
In conclusion, the AA model that we developed represents a useful tool for the identification and quantification of VF loss patterns in DOA beyond the MD. Although the VF machines were developed for the glaucoma analysis, the work presented here highlights the strength of the AA unsupervised ML method to retrieve and provide useful information also for the analysis of VF loss in optic neuropathies characterized by central scotoma. 
Future application of the DOA AA model may be further implemented by adding structural data (GCL), ultimately providing major insights on this disease. Furthermore, using the DOA AA model as the basis for longitudinal data analysis could be useful for progression analysis that may, for example, be implemented in the protocol of clinical trials as additional parameter to overcome the limits of classic outcome measure, as VA and MD, in defining a real functional improvement. 
Acknowledgments
Supported by a research scholarship funded by the European Union-Next Generation EU, Mission 4 Component 1 CUP J33C24001310009 to Catarina P. Coutinho. 
Ferdinando Zanchetta's research has been carried out under a research contract cofounded by the European Union and PON Ricerca e Innovazione 2014-2020 as in the art. 24, comma 3, lett. a), of the Legge 30 dicembre 2010, n. 240 e s.m.i. and of D.M. 10 agosto 2021 n. 1062. 
The research of Rita Fioresi and Ferdinando Zanchetta was supported by CaLISTA CA 21109, CaLIGOLA MSCA-2021-SE-01-101086123, MSCA-DN CaLiForNIA—101119552, PNRRMNESYS, the PNRR National Center for HPC, Big Data, and Quantum Computings SimQuSec; INFN Sezione Bologna, Gast initiative and GNSAGA Indam. 
Presented in part at the Association for Research in Vision and Ophthalmology 2023 Annual Meeting, New Orleans, Louisianna, USA, April 23–27, 2023. 
Disclosure: C.P. Coutinho, None; F. Zanchetta, None; M. Carbonelli, None; M. Battista, None; A. Galzignato, None; C. La Morgia, None; G. Amore, None; M. Romagnoli, None; G. Savini, None; L. Brotto, None; P. Nucci, None; L. Caporali, None; F. Bandello, None; V. Carelli, None; M.L. Cascavilla, None; R. Fioresi, None; P. Barboni, None 
References
Yu-Wai-Man P, Griffiths PG, Burke A, et al. The prevalence and natural history of dominant optic atrophy due to OPA1 mutations. Ophthalmology. 2010; 117(8): 1538–1546. [CrossRef] [PubMed]
Rocatcher A, Desquiret-Dumas V, Charif M, et al. The top 10 most frequently involved genes in hereditary optic neuropathies in 2186 probands. Brain. 2023; 146(2): 455–460. [CrossRef] [PubMed]
Fiorini C, Ormanbekova D, Palombo F, et al. The Italian reappraisal of the most frequent genetic defects in hereditary optic neuropathies and the global top 10. Brain. 2023; 146(9): 67–70. [CrossRef]
Lenaers G, Hamel C, Delettre C, et al. Dominant optic atrophy. Orphanet J Rare Dis. 2012; 7: 46. [CrossRef] [PubMed]
Ferré M, Bonneau D, Milea D. Molecular screening of 980 cases of suspected hereditary optic neuropathy with a report on 77 novel OPA1 mutations. Hum Mutat. 2009; 30(7): 692–697. [CrossRef]
Fuhrmann N, Alavi MV, Bitoun P, et al. Genomic rearrangements in OPA1 are frequent in patients with autosomal dominant optic atrophy. J Med Genet. 2009; 46(2): 136–144. [CrossRef] [PubMed]
Yu-Wai-Man P, Trenell MI, Hollingsworth KG, Griffiths PG, Chinnery PF. OPA1 mutations impair mitochondrial function in both pure and complicated dominant optic atrophy. Brain. 2011; 134(4): 164. [CrossRef]
Ham M, Han J, Osann K, Smith M, Kimonis V. Meta-analysis of genotype-phenotype analysis of OPA1 mutations in autosomal dominant optic atrophy. Mitochondrion. 2019; 46: 262–269. [CrossRef] [PubMed]
Kjer P . Hereditary infantile optic atrophy with dominant transmission. Acta Genet Stat Med. 1957; 7(2): 290–291. [PubMed]
Cohn AC, Toomes C, Hewitt AW, et al. The natural history of OPA1-related autosomal dominant optic atrophy. Br J Ophthalmol. 2008; 92(10): 1333–1336. [CrossRef] [PubMed]
Votruba M, Fitzke FW, Holder GE, Carter A, Bhattacharya SS, Moore AT. Clinical features in affected individuals from 21 pedigrees with dominant optic atrophy. Arch Ophthalmol. 1998; 116(3): 351–358. [CrossRef] [PubMed]
Kjer B, Eiberg H, Kjer P, Rosenberg T. Dominant optic atrophy mapped to chromosome 3q region. II. Clinical and epidemiological aspects. Acta Ophthalmol Scand. 1996; 74(1): 3–7. [CrossRef] [PubMed]
Lenaers G, Neutzner A, Le Dantec Y, et al. Dominant optic atrophy: culprit mitochondria in the optic nerve. Prog Retin Eye Res. 2021; 83: 100935. [CrossRef] [PubMed]
Manchester PT, Jr, Calhou FP, Jr. Dominant hereditary optic atrophy with bitemporal field defects. AMA Arch Ophthalmol. 1958; 60(3): 479–484. [CrossRef] [PubMed]
Battista M, Coutinho CP, Berni A, et al. Sectorial ganglion cell complex thickness as biomarker of vision outcome in patients with dominant optic atrophy. Invest Ophthalmol Vis Sci. 2024; 65(1): 24. [CrossRef] [PubMed]
Barboni P, Savini G, Cascavilla ML, et al. Early macular retinal ganglion cell loss in dominant optic atrophy: genotype-phenotype correlation. Am J Ophthalmol. 2014; 158(3): 628–636. [CrossRef] [PubMed]
Rönnbäck C, Milea D, Larsen M. Imaging of the macula indicates early completion of structural deficit in autosomal-dominant optic atrophy. Ophthalmology. 2013; 120(12): 2672–2677. [CrossRef] [PubMed]
Elze T, Pasquale LR, Shen LQ, Chen TC, Wiggs JL, Bex PJ. Patterns of functional vision loss in glaucoma determined with archetypal analysis. J R Soc Interface. 2015; 12(103): 20141118. [CrossRef] [PubMed]
Wang M, Shen LQ, Pasquale LR, et al. An artificial intelligence approach to detect visual field progression in glaucoma based on spatial pattern analysis. Invest Ophthalmol Vis Sci. 2019; 60(1): 365–375. [CrossRef] [PubMed]
Doshi H, Solli E, Elze T, Pasquale LR, Wall M, Kupersmith MJ. Unsupervised machine learning identifies quantifiable patterns of visual field loss in idiopathic intracranial hypertension. Transl Vis Sci Technol. 2021; 10(9): 37. [CrossRef] [PubMed]
Solli E, Doshi H, Elze T, Pasquale L, Wall M, Kupersmith M. Archetypal analysis reveals quantifiable patterns of visual field loss in optic neuritis. Transl Vis Sci Technol. 2022; 11(1): 27. [CrossRef] [PubMed]
Saifee M, Wu J, Liu Y, et al. Development and validation of automated visual field report extraction platform using computer vision tools. Front Med (Lausanne). 2021; 8: 625487. [CrossRef] [PubMed]
Motevalli Soumehsaraei B, Barnard A. Archetypal Analysis Package. CSIRO. v1. Software. 2019, https://doi.org/10.25919/5d3958889f7ff.
Cutler A, Breiman L. Archetypal analysis. Technometrics. 1994; 36(4): 338–347. [CrossRef]
Hastie T, Tibshirani R, Friedman JH. The elements of statistical learning. 2nd edition. New York, NY: Springer; 2017.
Thorndike RL . Who belongs in the family? Psychometrika. 1953; 18: 267–276. [CrossRef]
Syakur MA, Khotimah BK, Rochman EMS, Satoto BD. Integration K-means clustering method and elbow method for identification of the best customer profile cluster. IOP Conf Ser Mater Sci Eng. 2018; 335: 012017.
Kruger JM, Almer Z, Almog Y, et al. A consensus statement on the terminology for automated visual field abnormalities. J Neuroophthalmol. 2022; 42(4): 483–488. [CrossRef] [PubMed]
Keltner JL, Johnson CA, Cello KE, et al. Ocular Hypertension Treatment Study Group. Classification of visual field abnormalities in the ocular hypertension treatment study. Arch Ophthalmol. 2003; 121(5): 643–650. [CrossRef] [PubMed]
Figure 1.
 
Total deviation plot transformation of a visual field in format 30-2 (74 TD values, TD47 and TD57 represent the blind spot) to a 24-2 format (52 TD values) by excluding the 22 total deviation values represented in red.
Figure 1.
 
Total deviation plot transformation of a visual field in format 30-2 (74 TD values, TD47 and TD57 represent the blind spot) to a 24-2 format (52 TD values) by excluding the 22 total deviation values represented in red.
Figure 2.
 
(A) R2 scores change with the number of archetypes; (B) R2 scores difference against with the number of archetypes.
Figure 2.
 
(A) R2 scores change with the number of archetypes; (B) R2 scores difference against with the number of archetypes.
Figure 3.
 
The archetypes obtained with the respective relative weights in percentage. Colors are associated with the total deviations: blue = above normal; and red = below normal (visual field impairment). Points outside the VF are given a dark blue color.
Figure 3.
 
The archetypes obtained with the respective relative weights in percentage. Colors are associated with the total deviations: blue = above normal; and red = below normal (visual field impairment). Points outside the VF are given a dark blue color.
Figure 4.
 
Spearman correlation between the Composite AT Sum Score and the mean deviation (dB) on the left side and the visual acuity (logMAR) on the right side.
Figure 4.
 
Spearman correlation between the Composite AT Sum Score and the mean deviation (dB) on the left side and the visual acuity (logMAR) on the right side.
Figure 5.
 
Application of the 8-ATs model for the decomposition of: (A) a left eye 24-2 VF and (B) a 30-2 VF regarding the right eye. On the top left, the VF and on the bottom left the respective TDs. On the right are shown the used ATs and their respective RWs for the VF decomposition according to the developed model.
Figure 5.
 
Application of the 8-ATs model for the decomposition of: (A) a left eye 24-2 VF and (B) a 30-2 VF regarding the right eye. On the top left, the VF and on the bottom left the respective TDs. On the right are shown the used ATs and their respective RWs for the VF decomposition according to the developed model.
Table 1.
 
Demographic Characteristics of Study Cohort
Table 1.
 
Demographic Characteristics of Study Cohort
Table 2.
 
Archetype's Relative Weight in Percentage in the Developed Model (Training Set) and for the Decomposition of the VFs of the Test Set
Table 2.
 
Archetype's Relative Weight in Percentage in the Developed Model (Training Set) and for the Decomposition of the VFs of the Test Set
Table 3.
 
Spearman Correlation Coefficients Between the Archetype's Relative Weight and the Mean Deviation, in dB, and Visual Acuity, in logMAR
Table 3.
 
Spearman Correlation Coefficients Between the Archetype's Relative Weight and the Mean Deviation, in dB, and Visual Acuity, in logMAR
Table 4.
 
Archetype's Relative Weight in Percentage for the Haploinsufficiency, Missense Non-GTPase and Missense GTPase Visual Field Decomposition and Global and Post Hoc Test P Values for the Relative Weight Comparison
Table 4.
 
Archetype's Relative Weight in Percentage for the Haploinsufficiency, Missense Non-GTPase and Missense GTPase Visual Field Decomposition and Global and Post Hoc Test P Values for the Relative Weight Comparison
×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×