Artificial intelligence Integration in the diagnosis, prognosis and diabetic neovascular glaucoma treatment

Abstract

This work is focused on key aspects of the diagnosis, prognosis and treatment of neovascular glaucoma of diabetic origin based on machine learning approaches and, in particular, various architectures artificial neural models. An analysis of the relevance, priority provisions and advantages of using machine learning methods is carried out, the existing approaches used in modern literature in the context of the topic under study are considered, the specifics of their integration into the process of diagnostic analysis of the feature space of an aggregated and labeled by the authors data set on patients with visual problems are described, in particular, those suffering from neovascular glaucoma of diabetic origin. A correlation analysis of input features was carried out, 3 different models of artificial neural networks were built, trained and tested, metrics for assessing the accuracy of their work were experimentally calculated and studied, and statistical indicators were determined, including errors and losses, characterizing their generalizing ability. Analysis of the results obtained from the studies made it possible to identify the prevailing input features and evaluate their impact on the target output variable and the overall significance in the feature space of the data set, as well as to establish the most suitable models for data analysis in terms of their accuracy and speed. The conducted research made it possible to establish the fact of a greater degree deep learning artificial neural networks models fully connected adaptability for the analyzed data set.

Keywords

• Artificial intelligence
• neovascular glaucoma
• diagnosis eye treatment
• neural networks
• data analysis
• data mining
• machine learning 1

1. Introduction

and their combination are successfully analyzed and classified to determine the severity of the disease, its progression and/or referral for specialized treatment.

Diagnostic search algorithms for patients with glaucoma have undergone further changes, becoming more and more complex, and to improve existing methodologies it is necessary to use more reliable classification diseases signs. A significant contribution in this direction is made by modern methods and models for assessing and predicting the effectiveness of patient treatment, actively developed and implemented by specialists in the field of prevention and eye pathologies diagnosis. It should be noted that in the process of inflammation assessing indicators and intraocular circulation in patients collected during statistical data examination, an important aspect of the procedure for their analysis is indicators diagnostic significance [ 2 ]. This is relevant both for identifying and formalizing key factors that have the greatest impact on the visual organs integral state, and for developing and implementing preventive measures and developing protocols ophthalmic diseases effective treatment.

Due to the lack of clearly identified patterns and correlations in the feature space of aggregated vision organs as well as due to the high degree of uncertainty in mutual influence, carrying out the processes of assessing and analyzing data in a manual format is difficult and labor-intensive. Therefore, it is rational to use modern methods of data mining, in particular machine learning (ML) algorithms, in order to automate the process of searching for hidden patterns in data and assessing the level of feature space diagnostic significance, which is especially justified in the presence of large data volumes [ 3 ]. This allows us to provide solutions to various applied problems in the field of ophthalmology. In particular, it helps to increase the efficiency of the decision support process for assessing the condition and the use of targeted treatments for diabetic neovascular glaucoma [4] through the use of ML models to automate data research and search for individual signs correlations with each other. There are significant results from the use of ML in the analysis of such pathological eye conditions as diabetic retinopathy, agerelated macular degeneration, macular edema, glaucoma, cataracts, etc. In this case, color fundus photographs, images obtained during fluorescein angiography and autofluorescence, OCT scans, fields of view [ 5, 6, 7 ].

The defining characteristic of ML algorithms is the quality of their predictions, which improves with experience. The more data provided, the better the forecasting model. In this context, difficulties arise due to various factors, including [ 8 ]: • • • • • insufficiently high generalization ability of classical ML models; absence of memory effect in models; insufficient adaptability of models to data, taking into account the specifics under consideration; the need to compile ML models ensembles to ensure a higher degree of accuracy of the forecast values they generate; time costs for preliminary data scaling and preprocessing.

In this regard, a perspective way is to use more advanced, modern and efficient deep ML models, which are artificial neural networks (ANN) of various architectures, topologies and hierarchically interconnected.

An important aspect in the problem we are considering is the study of different ML models among themselves to experimentally identify an estimate of the most appropriate hyperparameters values, which affect the output metrics for assessing used models quality [ 9 ]. In this context, an approach based on deep learning and ANN makes it possible to more effectively solve classification and regression problems, taking into account the specifics of the application area under consideration, as well as to form various hybrid models that can have greater generalization ability and be more flexible in comparison with classical ML approaches, eliminate the imbalance of the feature space, perform the functions of data normalization and reduce the likelihood of overtraining models through the introduction of regularization methods [ 1 ]. An additional important aspect is the ability to minimize the risk of generating contradictory results from ANN models, for example, by integrating bagging and boosting techniques when creating model ensembles. In this regard, ANN using is more appropriate within the framework of the topic under consideration.

2. Analysis existing researches

In the medical field, there is an active use of various approaches, most often based on statistical methods and ML models for the analysis and evaluation of experimentally obtained results of diagnosis and various diseases treatment, including vision organs. Thus, the authors [ 10 ] adapt ML learning algorithms to solve the classification problem, achieving significant results of both high accuracy and completeness on the collected dataset of ophthalmological indicators. However, the data analyzed by the authors is often not complete and contains signs of synthesized subsamples, which complicates the procedure for assessing metric indicators. In [ 11 ], the authors use fully connected ANN models, comparing their work with each other and existing methods for classifying ML data, which is appropriate to demonstrate the capabilities and performance of these models in the context of their deployment in real diagnostic software and hardware tools for data analysis and evaluation.

The effectiveness of using ML was also established by the authors [ 12 ] in studies in the field of treating respiratory system evaluating methods when analyzing the significance of individual signs and diseases course intensity indicators.

According to a number of authors [ 13, 14 ], the use of methods for preventive intelligence medical data analysis in the context of assessing statistical indicators and correlations between individual data sets features allows the formation of an effective consistent data basis, thus organizing the process of automatic dimensionality reduction. As follows from the works [15, 16] on the use of ML in practice, the most effective are algorithms united in committees, this allows for models balancing and error values averaging for various metrics.

Thus, analyzing the research results in the reviewed authors works [ 10-16 ], it should be noted that more often they consider the possibility of using statistical approaches, as well as ML and ANN models, primarily to solve classification and regression problems in a heterogeneous feature space. However, outside the scope of research, questions remain related to assessing the individual medical indicators diagnostic significance, identifying correlations between input signs and assessing their specific weight values in the context of considering ophthalmological disease problems in patients with diabetic neovascular glaucoma.

The relevance of such studies is undoubted, due to the increasing disability of patients with this pathology throughout the world. In this regard, this work goal is to create and study different ANN models with generalization abilities and architecture for assessing the significance of features based on the collected data set on diagnostic indicators of inflammation and intraocular circulation in patients with diabetic neovascular glaucoma.

3. Collected data set description

The created set of the most significant target input and output features with medical interpretation was selected as a data set for analysis in applying ML process.

A retrospective cohort study of 127 patients (127 eyes) with a painful form of neovascular glaucoma (NVG) of diabetic origin was conducted at the Institute of Eye Diseases and Tissue Therapy named after. V.P. Filatov NAMS of Ukraine. The study protocol complied with the principles of the Declaration of Helsinki and was approved by the local bioethical committee. Written informed consent was obtained from all study participants. The average age of patients is 65.0 years (62-68).

The patients had uncontrolled intraocular pressure syndrome. The main criterion for assessing the success of treatment was control of IOP and achieving its reduction by 20% from the initial value after 12 months of observation and maintaining bestcorrected visual acuity (BCVA)

The preoperative visit (V0) was carried out on the eve of transscleral (TSC) cyclophotocoagulation (CPC), which was performed according to standard techniques. The laser power ranged from 850 to 1500 mW (Me 1100 mW), exposure time was 1.5-2 seconds, and laser coagulates average number was 22. In all patients, the need for repeat CFC TSC was determined at each postoperative visit. The criterion for repeated laser treatment was maintaining high IOP values. Preoperative laboratory parameters were determined: neutrophils (N), lymphocytes (L), platelets (Throm), monocytes (M), glycosylate rheoophthalmography.

The systemic immune inflammation index (SII = Throm×[N/L]) and systemic inflammation response index (SIRI = N×[M/L]) were calculated. The SIRI and SII scores were further divided into quartiles, and the RQ score into quintiles. The data has been imported, divided into input and output characteristics.

The target variable is the indicator which is a binary attribute. A fragment of the generated dataset is shown in Fig. 1. The results of related correlation analysis of the most significant features are shown in Fig. 2.

There is some imbalance in the output class values, which is not critical for research, and therefore the use of synthetic balancing methods is not advisable due to introducing additional noise risk into the data.

As part of the reconnaissance data analysis, a correlation analysis of the feature space was performed in order to identify obvious patterns between individual input variables. How can we note the indicators SIRI, SIRI quartiles. SII, SII quartiles have a high correlation with each other and with another input feature HbA1, which can introduce additional noise into the operation of models, and therefore it is necessary to take into account the nature of the influence of these features on the output target variable.

The correlation analysis showed a high positive relationship between the success of treatment and the indicators HbA1c (r = 0.751), SII (r = 0.876), SIRI quartiles (r = 0.874), IOP V0 (r = 0.611) and a high negative relationship with the RQ quintiles indicator (r=-0.807).

4. Development of artificial neural network models

We construct a computational process for constructing and studying ANN models to assess the accuracy of their training and classification. To do this, we use the Rapidminer system and blocks for importing a data set from a *.csv file, setting a target variable, and a subsystem for dividing the sample into test and validation. In order to conduct a systemic study and determine the ANN model type influence nature used on the final classification accuracy, assessed by different metrics, it is necessary to create several ANN models with different architectures. In this regard, we use 3 types of ANN models: single perceptron (SP), multilayer perceptron (MP) and deep multilayer neural network (DNN). The structure of the developed process for creating and modeling ANN operation is shown in the figure 3. All blocks are unitary commands aimed at processing data and performing processes for modeling the operation of ANN. importing the input data set using the Retrive block, formatting and converting them to a unified form for subsequent training of ANN models; setting input and output features for training the model by using the Set Role block; creation of a Validation container for combining blocks for constructing ANN models for the purpose of carrying out training processes, testing and evaluating the accuracy metrics (Apply model and Performance blocks);

ANN models research results export and visualization in tabular and graphical form.

Training and test data are divided among themselves in the proportion of 75% to 25%, respectively.

5. Experiments and results

10 computational experiments were carried out with different sets of parameters (hidden layers number, activation functions, learning rate coefficients, error rates). The average obtained metrics values for assessing accuracy are shown in the table below.

In particular, for a single perceptron, the value of the learning rate coefficient varied from 0.05 to 0.3. For a multilayer perceptron, 2 hidden layers of 5 and 3 neurons were created, respectively.

For the deep ANN, 50 hidden layers were created. The results of obtained ANN metrics evaluation to the first experiment are given in Table 1.

As we can see (Fig. 4), the least accurate classification results on both the training (about 84%) and training set were shown by the SP model (about 87%), which is due to its simplified architecture, the absence of all hidden layers and a general number of hyperparameters. affecting model final accuracy.

The MP and DNN models showed approximately the same results for metric evaluations on the training subset of data, however, the accuracy of DNN was 2.3% higher on the test subset, which may be due to a more complex model structure and a larger number of all performed computational iterations.

An additional experiment was aimed at reducing the dimension of input features. A unique identifier, calculated values of SIRI quartiles, SII, HbA1c, RQ quintiles, IOP V0 were used as input features; the output feature is

Research modelling was performed on the created three ANN models with the same parameters. The results of evaluating the metrics of the obtained NN models for the second experiment are given in Table 2, Fig. 5.

In the third experiment, gender was excluded from the original sample. This reduces noise effects because this feature does not carry a significant semantic load in the context of data analysis.

The results of evaluation of the metrics of the obtained NN models for the third experiment are given in Table 3, Fig. 6.

As a result of ANN constructing significant input parameters were established that most influence the effectiveness of treatment: SIRI quartiles, SII, HbA1c and RQ quintiles. During inflammation, leukocytes accumulate in the lesion and the speed of local blood flow decreases. Impaired retinal microcirculation is observed in patients with type 2 diabetes mellitus (T2DM) without clinically significant diabetic retinopathy [17].

Altered blood flow in patients with type 2 diabetes contributes to macrovascular (peripheral vascular disease and coronary artery disease) and microvascular (diabetic retinopathy and diabetic nephropathy) complications. Our results confirm that higher values of SIRI, SII, HbA1c and RQ indicate the severity of the disease and determine the need for diabetes mellitus stabilization [4], additional anti-inflammatory and anti-ischemic treatment in patients with painful NVG of diabetic origin to improve treatment prognosis.

6. Conclusion

The conducted research made it possible to establish a higher adaptability of deep learning models, in particular deep fully connected ANN models, for the analyzed data set. The SP model demonstrates unstable accuracy on the training and test samples in different experiments, which may be due to shortcomings in determining the weight values of features at different iterations, as well as limitations in generalization ability. The MP model is more stable in all experiments, demonstrating high accuracy and completeness; its construction speed is 25% higher than the SP model. Presumably, the accuracy of this model can be increased through a more efficient organization model hyperparameter values automating selection process. The DNN model is the most accurate, but its training process is the most resource-intensive and takes almost 4 times longer than the SP model. A promising direction for further research is the search for optimizing the ANN training process in order to minimize training and testing errors, as well as increase the generalization ability of models in general.

It has been established that the most significant diagnostic features of the collected dataset in the context of the problems under consideration are the indicator of glycosylated hemoglobin HbA1c of the volumetric intraocular circulation (RQ), the systemic inflammatory response index (SIRI) and the systemic inflammation index (SII). The data obtained allow us to conclude that the use of neural networks to predict the effectiveness of treatment is justified and timely. By taking into account and stabilizing blood sugar, as well as adjusting indicators of inflammation and microcirculation in patients with diabetic neovascular glaucoma, it is possible to ensure a timely significant reduction in intraocular pressure, maintain visual acuity, and therefore the quality of life of patients.

The data obtained allow us to conclude that the use of different ANN to predict the effectiveness of transcleral cyclophotocoagulation is justified. Timely prescribed treatment of identified changes in glycolyzed hemoglobin (HbA1c), volumetric intraocular circulation (RQ), systemic inflammatory response index (SIRI) and systemic inflammation index (SII) against the background of transcleral cyclophotocoagulation can provide a significant reduction in intraocular pressure of at least 20%, maintaining visual acuity and patients life quality with painful diabetic neovascular glaucoma. [15] E. Tokuç, M. Eksi, R. Kayar, S. Demir, R. Y. Bastug, M. Akyuz, M. Ozturk, Inflammation indexes and machine-learning algorithm in predicting urethroplasty success, Investigative and Clinical Urology, 65 (2024) 240. DOI: 10.4111/icu.20230302. [16] A. Abujaber, Y. Imam, I. Albalkhi, S. Yaseen, A. Nashwan, N. Akhtar, Utilizing machine learning to facilitate the early diagnosis of posterior circulation stroke, BMC Neurology, 24 (2024) 1-12.

DOI:10.1186/s12883-024-03638-8. [17] V. Cankurtaran, M. Inanc, K. Tekin, F. Turgut, Retinal Microcirculation in Predicting Diabetic Nephropathy in Type 2 Diabetic Patients without Retinopathy, Ophthalmologica, 243 4 (2020) 271-279. DOI: 10.1159/000504943.

References

[1] V. Vychuzhanin , N. Rudnichenko , Z. Sagova , M. Smieszek , V. V. Cherniavskyi , A. I. Golovan , M. V. Volodarets , Analysis and structuring diagnostic large volume data of technical condition of complex equipment in transport . IOP Conference Series: Materials Science and Engineering , Volume 776 , 24th Slovak-Polish International Scientific Conference on Machine Modelling and Simulations - MMS 2019 , 3-6 September 2019 , Liptovský Ján , Slovakia, 2019 pp. 1 - 11 . DOI: 10 .1088/ 1757 -899X/776/1/012049.

[2] N. Rudnichenko , V. Vychuzhanin , I. Petrov , D. Shibaev , Decision Support System for the Machine Learning Methods Selection in Big Data Mining , in: Proceedings of The Third International Workshop on Computer Modeling and Intelligent Systems (CMIS-2020) , CEURWS, 2608 , 2020 , pp. 872 - 885 .

[3] M. Alloghani , D. Al-Jumeily , J. Mustafina , A. Hussain , A.J. Aljaaf , A systematic review on supervised and unsupervised machine learning algorithms for data . Supervised and Unsupervised Learning for Data Science, Unsupervised and Semi-supervised Learning , Springer International Publishing, Cham ( 2014 ) 3 - 21 . DOI: 0.1007/978-3- 030 -22475- 2 _ 1 . The effect of the intercellular adhesion molecule-1 and glycated haemoglobin on the management of diabetic neovascular glaucoma , Romanian Journal of Ophthalmology, 68 2 ( 2024 ) 135 - 142 . DOI: 10 .22336/rjo. 2024 . 25 .

[5] J. De. Fauw , J.R. Ledsam , B. Romera-Paredes , Clinically applicable deep learning for diagnosis and referral in retinal disease , Nature Medicine, 24 9 ( 2018 ) 1342 1350 . DOI: 10 .1038/s41591- 018-0107-6.

[6] B. Al-Bander , W. Al-Nuaimy , B.M. Williams , Y. Zheng , Multiscale sequential convolutional neural networks for simultaneous detection of fovea and optic disc , Biomed Signal Process Control , 40 ( 2018 ) 91 101 . DOI: 10 .1016/j.bspc. 2017 . 09 .008.

[7] B.J. Antony , S. Maetschke , R. Garnavi , Automated summarisation of SDOCT volumes using deep learning: Transfer learning vs de novo trained networks , PLoS ONE , 14 5 ( 2019 ). DOI: 10 .1371/journal. pone. 0203726 .

[8] J. Bertomeu , Machine learning improves accounting: discussion, implementation and research opportunities, Review of Accounting Studies, 25 3 ( 2020 ) 1135 - 1155 . DOI: 10 .1007/s11142-020- 09554-9

[9] M. Solanki , Ms. Sharma, A Review of Data Mining Techniques and Its Applications , International Journal of Innovative Research in Computer Science & Technology , 122 ( 2021 ) 100 - 104 . DOI: 10 .55524/ijircst. 2021 . 9 .6.23.

[10] S. Sheshkal , M. Gundersen , M. Riegler , Ø. Utheim, K. Gundersen , H. Hammer , Classifying Dry Eye Disease Patients from Healthy Controls Using Machine Learning and Metabolomics Data ( 2024 ).

[11] A. Abdullah , A. Aldhahab , H. Abboodi , Review of Eye Diseases Detection and Classification Using Deep Learning Techniques , BIO Web of Conferences , 97 ( 2024 ). DOI: 10 .1051/bioconf/20249700012.

[12] A. Selim , M. Barz , O. Bhatti , H. Alam Tusfiqur , D. Sonntag , A review of machine learning in scanpath analysis for passive gaze-based interaction , Frontiers in Artificial Intelligence , 7 ( 2024 ). DOI: 10 .3389/frai. 2024 . 1391745 .

[13] J. Van Eijgen , V. Schuhmann , E.-L. Fingerroos , M. Renier , H. Burchert , J. Kröpfl , A. Craenenbroeck , V. Cornelissen , K. Gugleta , I. Stalmans, H. Hanssen , High-intensity interval training in patients with glaucoma (HIT-GLAUCOMA): protocol for a multicenter randomized controlled exercise trial , Frontiers in Physiology, 15 ( 2024 ). DOI:1349313. 10 .3389/fphys. 2024 . 1349313 .

[14] C.-H. Moon , Ah-Jun Lee , H.-Y. Jeon , E.- B. Kim , K.-S. Ha, Therapeutic effect of ultra-long-lasting human C-peptide delivery against hyperglycemia-induced neovascularization in diabetic retinopathy , Theranostics , 13 ( 2023 ) 2424 - 2438 . DOI: 10 .7150/thno.81714.