Machine Learning Models For Dry Eyes Diagnosis And Treatment Using Real World Clinical Data

Purpose:

To develop reliable Machine Learning (ML) models capable of using demographic and real-world clinical data to predict dry eye severity and dry eye type .

Real-world clinical data captured over a span of 18 months from 495 physicians.

Methods:

Real-world clinical data captured over span of 18 months were used to develop two machine learning diagnostic models for severity and type of dry eye. The acquired data had 25118 well-structured samples, where each sample had dozens of features and was annotated by domain experts. 495 physicians currently contributed 21190 total questionnaires. Correlation feature selection (CFS) was used to eliminate redundant and irrelevant features. Using stratified 10-fold cross-validation was shown to be the most effective ML technique. Support Vector Machines (SVM) was the best technique to fit our data. We developed two SVM models, namely SM (severity model) and TM (type model). The two models used the exact feature for prediction.

Results:

The SM model was successful in predicting different dry eye severity cases at AUC−ROC of 0.79 and AUC−PR of 0.61. Furthermore, the TM model was successful in predicting different dry eye type cases at AUC−ROC of 0.91 and AUC−PR of 0.94. We also verified the robustness of both SM and TM models by comparing their performance with nine baseline machine learning methods. The two proposed models outperformed all baseline methods based on AUC−ROC and AUC−PR. We will also discuss the treatment side to the model using DEWS-II models.

Conclusion:

As a validated dry eye prediction model, the ML models proposed in this study accurately predicted dry eye severity and type. Such models may improve health outcomes and provide early alerting to prevent the progression of dry eye disease. Future data will enhance these models and help us enhance treatment models and treatments based on dry eye prediction.