Near viewing behaviors predict educational system in a machine learning model

Abstract

Intensive education systems are believed to contribute to high rates of myopia. This study examined whether near-viewing behaviors in college students differ based on their pre-college educational systems and whether these behaviors can be used to classify students’ educational background using machine learning. Male students ages 18–33 years who attended either an intensive (ultra-Orthodox) or a standard school system (non-ultra-Orthodox) prior to college were recruited. Refractive error was measured and near-viewing behaviors were assessed using a wearable sensor during academic study periods. Compared to standard school students, intensive school students had significantly more myopic refraction (P < 0.03), spent more time viewing very near distances (P < 0.004) and less time viewing intermediate distances (P < 0.008) and had shorter near-viewing distances (P < 0.0001). Machine learning identified far-viewing episodes > 5 min and viewing distance during near-viewing as predictors of educational background.These findings suggest that educational environments are associated with distinct visual behavior patterns that may be linked to refractive development. The ability to use machine learning to predict educational systems based solely on near-viewing behaviors underscores its potential as a tool for investigating educational and behavioral factors and refractive outcomes.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-10108-9.

Introduction

Myopia is recognized as a global issue that poses significant public health concerns. The prevalence of myopia is increasing worldwide, with estimates suggesting that by 2050, half of the world’s population will be affected¹. In addition, the prevalence of high myopia is also increasing². Consequently, myopia-related ocular complications, such as glaucoma^3,4 cataracts^5,6 retinal detachment, and atrophy⁷ are also expected to rise. This trend highlights the need for a better understanding of factors contributing to myopia onset and progression.

Multiple factors, including genetic, environmental, and behavioral influences, are believed to play a role in the development and progression of myopia^8–15. Among these, educational demand, as highlighted by the International Myopia Institute, has emerged as a significant risk factor¹⁰. However, the specific aspects of education that contribute to myopia are not fully understood. Evidence suggests that near-viewing behaviors, including near distances shorter than 20–30 cm or continuous near work for over 30–45 min without a break^16,17 may impact myopia development more than the total duration spent on near work^18,19.

If near work behavior is a critical factor, comparing different educational systems with varying myopia prevalence may offer insights into which specific behaviors contribute to myopia development. The Jewish population in Israel presents a unique opportunity to investigate this, as the distinct educational systems, including ultra-Orthodox and, broadly, non-ultra-Orthodox, have been shown to correspond to differences in myopia prevalence¹⁴. Therefore, these populations may offer a model to understand the different rates of myopia in other populations with intense educational demands. For example, the educational system for ultra-Orthodox boys is similar to that found in urbanized East Asian countries, which involves intense near-work activities and long school days, beginning at the age of three years²⁰. In line with that, the prevalence of myopia in ultra-Orthodox males (80%) is also similar to that in urbanized East Asian countries²¹. On the other hand, non-ultra-Orthodox boys only begin formal education at the age of six and follow a standard curriculum years, similar to Western populations²². Furthermore, non-ultra-Orthodox males have a myopia prevalence of 30–50%, which is similar to the global average²³. Given these observations, Israel’s unique educational environment may offer broader insights into the relationship between myopia prevalence and educational systems, which will also be relavent to patterns seen in East Asian and Western countries.

Most prior studies assessing near work depend on subjective self- or parent-reports collected through questionnaires or diaries¹⁸. Recently, studies have utilized objective wearable sensors, such as the RangeLife²⁴ a novel near work analyzer²⁵ and the Clouclip^26–30 to collect objective data on near work duration and distance. These studies have shown that objective measures differ significantly compared to parental self-reported data^29,31,32 and offer a more comprehensive quantification of near viewing behaviors. However, wearable sensors generate large datasets that present new analytical challenges. Effectively processing and interpreting these data requires a multidisciplinary approach, integrating expertise from vision science, data science, and artificial intelligence.

The current study aimed to objectively assess the differences in near-viewing behaviors among college students who had attended either an intensive school system or a more traditional standard school system during their pre-college education. Previous research has demonstrated that male graduates of intensive school systems exhibit closer near-work distances than their counterparts from standard school systems when objectively measured during controlled tasks in a laboratory³⁰. Based on these findings, we hypothesized that viewing behaviors, specifically closer working distances and longer durations of near work, will be observed in graduates of intensive school systems. This study serves as a pilot investigation into the application of machine learning for identifying distinct near-viewing behaviors associated with different educational backgrounds. By demonstrating the feasibility of this approach, our findings provide a basis for future research exploring links between education and myopia using objective behavioral data.

Materials and methods

Participants

Healthy Jewish male college students in Israel between the ages of 18–33 years were recruited through advertisements. Those who responded and met the inclusion criteria were enrolled in the study. Each participant was provided with a verbal explanation of the study’s procedures and subsequently provided written informed consent, in compliance with the ethical guidelines approved by the Hadassah Academic College Ethics Committee. Inclusion criteria were visual acuity of 6/9 or better and no history of ocular disease or surgery, binocular vision anomalies, or systemic diseases known to affect the eyes. Participants were categorized as graduates of either an intensive (ultra-Orthodox) or standard (non-ultra-Orthodox) school system based on self-reporting and will be refered hereon as intensive school students and standard school students³³.

Clinical examination

Each participant underwent assessments for visual acuity (Snellen Chart) and cover test at both distance and near. Non-cycloplegic autorefraction was performed using a Visionix VX130 (Visionix, Luneau, Chartres, France). Three measurements were recorded in each eye, and spherical equivalent refraction was calculated. Participants who arrived without their spectacles or exhibited poor visual acuity with their usual correction were excluded and received a referral for a comprehensive eye examination.

Myopia was defined as spherical equivalent refraction of ≤−0.50 D in both eyes³⁴ and hyperopia was defined as spherical equivalent refraction of ≥ + 1.25 D in both eyes³⁰. High myopia was categorized as ≤−6.00 D³⁴. Participants with spherical equivalent refraction between >−0.50 D and < + 1.25 D in both eyes were classified as emmetropic.

Objective measurements

The Clouclip was used to continuously and objectively measure viewing distances. The Clouclip is a small, lightweight Bluetooth device that measures distance every five seconds using an infrared beam²⁶. The sensor has a reported field of view of 25 degrees and a validated measurement range of 5–120 cm²⁸. The Clouclip is equipped with angular acceleration sensors and, if no movement is detected for 40 s, it enters sleep mode for a minimum of two minutes or until movement is detected^28,35. Data collected by the Clouclip are stored in the device and later uploaded to a cloud server via a Bluetooth connection and a smartphone application.

The Clouclip was mounted on the right temple of a spectacle frame, manually adding an additional tilt of approximate 10° downward to reliably capture the viewing targets, as described previously²⁸. For participants who wore prescription glasses, the Clouclip was mounted on their habitual frames, while for those without prescriptions, the Clouclip was mounted on a provided frame with plano lenses. Participants were instructed to charge the Clouclip overnight before use and to wear it for at least one full day of classroom-based academic study. Some participants wore the device up to 6 days, with an average of 1.9 days.

Data analysis

Objective behavior measurements: cleaning procedures

Data from the Clouclip were downloaded from the cloud to Excel (Microsoft.com). Each file included time, distance (centimeters), and illumination (lux) in 5 s epochs. To prepare the data for analysis, the first step involved removing data indicative of the Clouclip being in sleep mode, which was identified as distance readings of −1 cm and is indicative of the participant removing the glasses. In the second step, since the study focused on indoor behavior, measurements with lux values greater than 1000 (indicating outdoor exposure)^8,36–40 were excluded. Finally, artifacts were removed. Artifacts were defined as groups of at least 300 consecutive measurements with distances less than 100 cm, where most of the measurements had distances below 10 cm, and both the mean and standard deviation of these distances were under 10 cm. These thresholds were determined through visual inspection of the data and best captured artifacts present in the dataset. Any measurements with distances below 10 cm that did not meet this criterion were retained. Following the data cleaning process, only data from participants with a minimum of one hour of valid data were included in the analysis.

Definition

Following data cleaning, the remaining data were categorized and analyzed under four main categories: (1) duration of near and far viewing, (2) viewing distance during near viewing, (3) continuous episodes of near and far viewing, and (4) illuminance. Time spent in near viewing included measurements with a distance between 10 and < 100 cm. Near viewing was further subdivided into very near viewing (10 to < 40 cm) and intermediate near viewing (40 to < 100 cm). Far viewing was defined as measurements outside of the near viewing range, specifically between 0 and < 10 cm and ≥ 100 cm. Although distances below 10 cm are technically closer than the near range, they were categorized as far viewing because they are not feasible and likely reflect Clouclip measurement error, as such distances were not observed during direct laboratory observations³⁰. Viewing distance during near viewing was recorded in cm for measurements within the range of near viewing (10 and < 100 cm), very near viewing (10 and < 40 cm), and intermediate near viewing (40 and < 100 cm). Illuminance in lux was averaged for near viewing (10 and < 100 cm) and for far viewing (0 and < 10 cm and ≥ 100 cm).

For the analysis of continuous episodes, we defined two viewing distance categories: near viewing (10 and < 100 cm) and far viewing (0 and < 10 cm and ≥ 100 cm). A continuous episode was defined as an uninterrupted sequence of measurements within the same viewing distance category (either near viewing or far viewing). The duration of near viewing and far viewing continuous episodes was classified into four groups: episodes lasting < 1 min, 1 to < 2 min, 2 to < 5 min, and more than 5 min. Two metrics were calculated for both near viewing and far viewing episodes: (1) the percentage of total time spent in episodes of different lengths within each viewing distance category, and (2) the percentage of episodes of different lengths relative to the total number of episodes in each category. Additionally, for comparative purposes, we implemented the Bhandari & Ostrin Model of continuous near viewing⁴¹. Briefly, this model distinguishes between two types of continuous near viewing: (1) Short-duration episodes of distances of 10 and < 100 cm for more than 1 min and up to 30 min, with far viewing not exceeding 20 s (whether scattered or consecutive), and (2) long-duration episodes of distances of 10 and < 100 cm for more than 30 min, with far viewing not exceeding 60 s (whether scattered or consecutive). Further details on the model are provided in Appendix 1, Table A1.

Statistical testing

Statistical analysis was conducted using Python’s Scipy scientific library version. Data are presented for the entire sample, as well as stratified by group (intensive and standard school). Summary statistics were calculated for both the overall sample and for each group. For numerical variables, means, standard deviations, and minimum and maximum values are reported. For categorical variables, percentages are presented.

Group differences were tested by first assessing normality using the Shapiro–Wilk test. For variables with a normal distribution, the parametric Student’s t-test was applied, while for non-normally distributed variables, the nonparametric Mann–Whitney U test was used. Categorical outcomes, such as myopia prevalence, were compared using a chi-squared test. Statistical significance was defined as p < 0.05.

The Pearson correlation coefficient was used to measure the relationships between spherical equivalent refraction and the following variables: (1) near viewing distance, (2) very near viewing distance, (3) intermediate near viewing distance, (4) time spent in near viewing, (5) time spent in very near viewing, and (6) time spent in intermediate near viewing. Outliers were identified using the Z-score test, with a threshold of 3 in absolute value, and any detected outliers were excluded from the analysis prior to applying Pearson correlation.

Machine learning

To predict the likelihood that an individual graduated from an intensive school system based solely on behavioral data, we evaluated several machine learning models (Appendix 2, Table A2). The process of model selection, feature handling, model evaluation, and the interpretation of results is described below.

We explored a diverse array of machine learning models to enhance prediction accuracy and robustness. We employed ensemble methods like Random Forest⁴² and Gradient Boosting⁴³ to improve generalizability and reduce overfitting. Random Forest utilizes multiple decision trees to produce stable, reliable predictions, whereas Gradient Boosting incrementally corrects errors using a series of weak learners. Logistic Regression⁴⁴ was also included for its straightforward effectiveness in binary classification tasks. To mitigate multicollinearity, which can skew model accuracy, we included also Lasso Regression⁴⁵ which uses L1 regularization to eliminate non-essential features by zeroing out their coefficients. Additionally, Naive Bayes was selected for its efficiency in generating probabilistic predictions in binary classification contexts.

Feature selection and multicollinearity management

Feature selection was critical in developing our predictive model, initially incorporating only those features that demonstrated statistical significance in statistical analyses. To proactively manage multicollinearity, we employed the Variance Inflation Factor (VIF)^46,47 ensuring that all features selected for the model maintained a VIF score of less than 10. Furthermore, we rigorously checked for pairwise correlations among features, excluding any that exhibited significant correlations (p-value < 0.05) to ensure the independence of the predictors.

Model evaluation

The performance of each model was evaluated using a stratified 5-fold cross-validation method⁴⁸. This approach not only ensures that each fold is a representative of the entire dataset by maintaining the same proportion of classes but also enhances the robustness of our evaluation against overfitting. We assessed several key performance metrics including accuracy, precision, recall, F1-score, and ROC AUC. The mean and standard deviation of these metrics were computed across the cross-validation folds to provide insights into the models’ consistency and reliability. Additionally, by comparing mean scores from both training and testing phases for each metric, we were able to detect potential overfitting, ensuring that the selected models perform reliably under various scenarios.

Model performance and selection

The selection of the final model was guided by the evaluation of the performance metrics outlined previously. We prioritized models that not only achieved high performance across metrics such as accuracy, precision, recall, F1-score, and ROC AUC, but also exhibited minimal discrepancies between training and testing phases. This careful balance ensured that the chosen models were robust and reliable, demonstrating their ability to generalize well to new datasets.

SHAP analysis

The interpretation of model predictions was conducted using SHapley Additive exPlanations (SHAP) values⁴⁹. SHAP provides a consistent and interpretable framework for assessing feature importance by assigning an individual contribution value to each feature. This allows for a more transparent understanding of how behavioral features influenced the likelihood of an individual being classified as having attended the ultra-Orthodox educational system.

Results

Participants

Demographic and clinical findings for participants are shown in Table 1. Participants included 79 young adult males (43 intensive school students and 36 standard school students). Data from 15 participants were excluded due to Clouclip malfunctions (8 intensive school students and 7 standard school students). Therefore, data from 64 participants, ages 18–33 years (mean 24.3 ± 4.0 years, 28 instensive school and 36 standard school students) were included in the analyses. Forty-six participants wore habitual spectacles (22 instensive school students, 24 standard school students), all of which were single vision lenses. Remaining participants were provided with plano lens spectacles.

Table 1.

Demographic and clinical data (mean ± SD and range) overall (N = 64) and by group, intensive school (N = 28) and standard school (N = 36). *Significant differences between groups.

	All	Intensive School	Standard School	P Value
Number of participants	64	28	36
Age (range), (years)	24.2 ± 4.0 (18 to 33)	22.0 ± 4.2 (18 to 33)	26.0 ± 2.8 (20 to 33)	< 0.001*
Myopia Prevalence (N)	75% (48)	82.1% (23)	69.4% (25)	0.24
High Myopia Prevalence (N)	19% (12)	32.1% (9)	8.3% (3)	0.02*
Spherical Equivalent Refraction (range), (D)	−2.86 ± 3.21 (−11.37 to + 1.04)	−4.05 ± 3.75 (−11.37 to + 1.04)	−1.94 ± 2.40 (−9.80 to + 0.81)	0.02*

The intensive school group was significantly younger than the standard school group (22.0 ± 4.2 vs. 26.0 ± 2.8 years, respectively; U = 212.5, P < 0.0001). There was no significant difference in the prevalence of myopia between the groups (X² = 1.35, P = 0.24). However, the intensive school group had significantly more myopic refraction than the standard school group (−4.05 ± 3.75 vs. −1.94 ± 2.40 D, respectively; U = 336, P = 0.02). In addition, the intensive school group had a higher prevalence of high myopia than the standard school group (32% vs. 8.3%, respectively; X² = 5.86, P = 0.02).

Objective behavior measurements: data cleaning

A representative example of data cleaning for one participant is shown in Fig. 1. The percentage of data retained at each step of the data cleaning and preparation process is detailed as follows. First, sleep data (i.e. when the Clouclip went into sleep mode) were removed, resulting in an average reduction of 10.6% of the total measurements. Next, measurements taken outdoors (with illuminance > 1000 lx) were excluded, which led to the removal of an additional 6.4% of the remaining data. Finally, artifacts were identified and removed, further reducing the dataset by 11.5%. Altogether 26% of the data was cleaned, with the remaining 73.9% included in the final analysis. There were no significant differences between groups in the proportion of data removed due to artifacts.

Figure 1.

Raw distance data for one representative participant. Data identified as collected while the device was in sleep mode are shown in green (distance = −1). Data collected outdoors are shown in orange (lux ≥ 1000). Artifacts are shown in red. The blue symbols are the remaining data that was used in the analysis. Points with viewing distances of 204 cm represent measurements beyond the maximum range that the Clouclip device can accurately measure and are considered far distance viewing.

Results of objective behavior measurements

Time in near and Far viewing

Metrics of visual behavior for intensive school and standard school groups are shown in Table 2. The percentage of time spent at near viewing (10 to < 100 cm) was comparable between the intensive school and standard school groups (56.6 ± 13.9% vs. 54.1 ± 15.3%, respectively, T = 0.68, P = 0.50). However, the intensive school group spent significantly more time at very near viewing (10 to < 40 cm) compared to the standard school group (31.2 ± 15.8% vs. 20.1 ± 11.3%, respectively, U = 718, P = 0.004). Conversely, the intensive school group spent significantly less time at intermediate near viewing (40 to < 100 cm) than the standard school group (25.4 ± 12.5% vs. 34.0 ± 13.1%, respectively, U = 307, P = 0.008). There were no significant differences between groups in the percentage of time spent in far viewing (0 to < 10 cm and ≥ 100 cm) (43.4 ± 13.9% vs. 45.9 ± 15.3%, respectively, T = −0.68, P = 0.11).

Table 2.

Percent of time spent at near and Far viewing (mean ± SD and range) overall (N = 64) and by group, intensive school (N = 28) and standard school (N = 36). *Significant differences between groups.

	All	Intensive School	Standard School	P Value
Percent of time in near viewing (10 to < 100 cm)	55.2 ± 14.6 (19.6 to 87.8)	56.6 ± 13.9 (22.9 to 79.8)	54.1 ± 15.3 (19.6 to 87.8)	0.50
Percent of time in very near viewing (10 to < 40 cm)	24.9 ± 14.5 (4.4 to 65.7)	31.2 ± 15.8 (5.2 to 65.7)	20.1 ± 11.3 (4.4 to 51.8)	0.004*
Percent of time at intermediate near viewing (40 to < 100 cm)	30.3 ± 13.4 (8.3 to 57.3)	25.4 ± 12.5 (8.9 to 53.8)	34.0 ± 13.1 (8.3 to 57.3)	0.008*
Percent of time in far viewing (0 to < 100 and ≥ 100 cm)	44.8 ± 14.6 (12.2 to 80.5)	43.4 ± 13.9 (20.2 to 77.1)	45.9 ± 15.3 (12.2 to 80.5)	0.11

Viewing distances

Mean viewing distances are shown in Table 3. The intensive school group maintained significantly shorter viewing distances during near viewing (10 to < 100 cm) compared to the standard school group (41.6 ± 7.9 cm vs. 48.4 ± 7.7 cm, respectively; T = −3.47, P < 0.001). Similarly, mean viewing distance during very near viewing (10 to < 40 cm) was significantly less in the intensive school group compared to the standard school group (27.0 ± 2.8 cm vs. 28.7 ± 2.5 cm, respectively; T = −2.61, P = 0.01). However, no significant difference was observed in the intermediate near viewing distance (40 to < 100 cm) between groups (58.0 ± 4.9 cm vs. 59.9 ± 4.9 cm, respectively; T = −1.55, P = 0.13).

Table 3.

Near viewing distances (mean cm ± sd and range) overall (N = 64) and by group, intensive school (N = 28) and standard school (N = 36). *Significant differences between groups.

	All	Intensive school	Standard School	P Value
Viewing distance (cm) during near viewing (10 to < 100 cm)	45.4 ± 8.5 29.2 to 68.8	41.6 ± 7.9 29.18 to 57.75	48.4 ± 7.7 34.2 to 68.8	< 0.001*
Viewing distance (cm) during very near viewing (10 to < 40 cm)	28.0 ± 2.8 19.1 to 33.6	27.0 ± 2.8 19.1 to 32.3	28.7 ± 2.5 23.1 to 33.6	0.01*
Viewing distance (cm) during intermediate near viewing (40 to < 100 cm)	59.1 ± 4.9 49.6 to 75.3	58.0 ± 4.9 49.6 to 67.9	59.9 ± 4.9 52.6 to 75.3	0.13

When examining the correlation between the distance of near viewing (10 to < 100 cm) and spherical equivalent refraction, no significant correlation was found (R = 0.23, P = 0.06, N = 64). One participant was identified as an outlier for very near viewing distance (10 to < 40 cm) using the Z-score test, with a Z-score exceeding the threshold of 3. After excluding the outlier, the correlation between spherical equivalent refraction and very near viewing (10 to < 40 cm) remained statistically insignificant (R = 0.24, P = 0.06, N = 63). Similarly, no significant correlations were observed between spherical equivalent refraction and (1) intermediate work viewing distance (40 to < 100 cm; R = −0.11, P = 0.37, N = 63), (2) duration of near viewing (10 to < 100 cm; R = −0.03, P = 0.79, N = 64), (3) duration of very near viewing (10 to < 40 cm; R = −0.20, P = 0.11, N = 64), or (4) duration of intermediate near viewing (40 to < 100 cm; R = 0.17, P = 0.15, N = 64).

Continuous episodes

Continuous episodes of each viewing category were quantified by group (Table 4). There were no significant differences observed in continuous episodes of near viewing between the intensive school and standard school groups. However, significant differences emerged when examining far viewing episodes (Fig. 2). The intensive school group exhibited a higher percentage of far viewing episodes lasting less than one minute compared to the standard school group (92.1 ± 4.7% vs. 89.3 ± 5.1%, U = 708, P = 0.006), as well as a significantly greater percentage of time spent in this far viewing episodes (57.1 ± 15.1% vs. 45.6 ± 15.5%, T = 2.96, P = 0.004). For far viewing episodes lasting 1–2 min, the intensive school group had significantly fewer episodes (5.0 ± 2.9% vs. 6.1 ± 2.2%, U = 327, P = 0.02), although the percentage of time spent during these episodes was similar between the groups (17.1 ± 5.7% vs. 17.4 ± 5.7%, U = 490, P = 0.85). Regarding far viewing episodes lasting 2–5 min, the number of episodes was less in the intensive school group (2.4 ± 2.1% vs. 3.3 ± 2.2%, U = 336, P = 0.02), but no significant differences were found in the percentage of time spent during these episodes (16.1 ± 9.7% vs. 19.2 ± 6.5%, T = −1.50, P = 0.14). Lastly, for far viewing episodes longer than 5 min, the intensive school group had significantly fewer far viewing episodes (0.5 ± 0.7% vs. 1.3 ± 1.6%, U = 319, P = 0.01), with a correspondingly lower percentage of time spent during these episodes (9.7 ± 10.5% vs. 17.6 ± 15.5%, U = 340, P = 0.03).

Table 4.

Near and Far viewing continuous episodes (mean ± SD and range) overall (N = 64) and by group, intensive school (N = 28) and standard school (N = 36). *Significant differences between groups.

	All	Intensive school	Standard School	p
% Near viewing episodes < 1 min (out of total near viewing episodes)	86.8 ± 7.5 62.5 to 98.7	87.2 ± 7.6 67.4 to 98.7	86.5 ± 7.5 62.5 to 95.1	0.81
% Time in near viewing episodes < 1 min (out of total time in near viewing episodes)	41.7 ± 17.3 2.4 to 91.1	43.3 ± 17.3 9.7 to 91.1	40.4 ± 17.4 2.4 to 73.1	0.51
% Near viewing episodes 1–2 min (out of total near viewing episodes)	6.9 ± 3.1 0.0 to 15.6	6.9 ± 3.3 1.3 to 15.6	6.9 ± 2.9 0.0 to 14.9	0.95
% Time in near viewing episodes 1–2 min (out of total time spent in near viewing episodes)	17.0 ± 5.5 0.0 to 28.6	17.1 ± 4.5 8.3 to 28.2	16.9 ± 6.2 0.0 to 28.6	0.88
% Near viewing episodes 2–5 min (out of total near viewing episodes)	4.2 ± 2.6 0.0 to 11.6	4.3 ± 2.8 0.0 to 11.6	4.1 ± 2.5 0.7 to 9.7	0.88
% Time in near viewing episodes 2–5 min (out of total time spent in near viewing episodes	20.3 ± 7.2 0.0 to 35.9	22.0 ± 7.8 0.0 to 35.9	19.0 ± 6.6 5.3 to 31.9	0.10
% Near viewing episodes > 5 min (out of total near viewing episodes)	2.1 ± 3.7 0 to 23.5	1.6 ± 2.0 0.0 to 8.1	2.5 ± 4.6 0.0 to 23.5	0.51
% Time in near viewing episodes > 5 min (out of total time spent in near viewing episodes)	21.0 ± 19.0 0.0 to 89.3	17.6 ± 14.4 0.0 to 53.8	23.7 ± 21.8 0.0 to 89.3	0.33
% Far viewing episodes < 1 min (out of total far viewing episodes)	90.5 ± 5.1 75.3 to 97.8	92.1 ± 4.7 79.0 to 97.3	89.3 ± 5.1 75.3 to 97.8	0.006*
% Time in far viewing episodes < 1 min (out of total time spent in far viewing episodes)	50.7 ± 16.3 11.7 to 85.1	57.1 ± 15.1 31.5 to 85.1	45.6 ± 15.5 11.7 to 75.6	0.004*
% Far viewing episodes 1–2 min (out of total far viewing episodes)	5.7 ± 2.6 0.0 to 12.4	5.0 ± 2.9 0.0 to 12.4	6.1 ± 2.2 1.1 to 10.1	0.02*
% Time in far viewing episodes 1–2 min (out of total time spent in far viewing episodes)	17.4 ± 5.7 0.0 to 30.6	17.1 ± 5.7 0.0 to 29.7	17.6 ± 5.7 6.8 to 30.6	0.85
% Far viewing episodes 2–5 min (out of total far viewing episodes)	2.9 ± 2.2 0.0 to 11.9	2.4 ± 2.1 0.0 to 8.6	3.3 ± 2.2 0.7 to 11.9	0.02*
% Time in far viewing episodes 2–5 min (out of total time spent in far viewing episodes)	17.9 ± 8.1 0.0 to 38.8	16.1 ± 9.7 0.0 to 38.8	19.2 ± 6.5 7.1 to 33.4	0.14
% Far viewing episodes > 5 min (out of total far viewing episodes)	0.9 ± 1.3 0.0 to 5.7	0.5 ± 0.7 0.0 to 3.2	1.3 ± 1.6 0.0 to 5.7	0.01*
% Time in far viewing episodes > 5 min (out of total time spent in far viewing episodes)	14.2 ± 14.0 0.0 to 58.6	9.7 ± 10.5 0.0 to 45.3	17.6 ± 15.5 0.0 to 58.6	0.03*

Figure 2.

(A) Percentage of far viewing episodes of different lengths and (B) percentage of time spent in far viewing episodes of different lengths. Black bars represent the intensive school group, and open bars represent the standard school group. Error bars indicate the standard error of the mean. *Significant differences between groups.

Illuminance

Illuminance did not differ significantly between the two groups (Table 5). The intensive school group had similar illuminance levels during near viewing compared to the standard school group (115.8 ± 68.6 lx vs. 100.3 ± 41.9 lx, respectively, U = 542, P = 0.61). Similarly, illuminance during far viewing was comparable between the groups (132.2 ± 93.0 lx vs. 108.8 ± 40.3 lx, respectively, U = 523, P = 0.80).

Table 5.

Illuminance (mean ± SD and range) overall (N = 64) and by group, intensive school (N = 28) and standard school (N = 36).

	All	Intensive school	Standard School	P Value
Illuminance (lux) during near viewing (10 and < 100 cm)	107.1 ± 55.3 (28.6 to 314.6)	115.8 ± 68.6 (28.6 to 314. 6)	100.3 ± 41.9 (40.8 to 193.4)	0.61
Illuminance (lux) during far viewing (0 and < 10 and ≥ 100 cm)	119.0 ± 68.9 (15.1 to 378.7)	132.2 ± 93.0 (18.4 to 378.7)	108.8 ± 40.3 (15.1 to 194.1)	0.80

Machine learning for predicting educational system

As described in the methods section, our rigorous feature selection led to the identification of two pivotal behavioral features: the viewing distance during near viewing (10 to < 100 cm) and the percentage of far viewing episodes longer than 5 min out of the total far viewing. These features were selected due to their statistical significance and minimal inter-correlation, meeting the criteria of having a Variance Inflation Factor (VIF) below 10 and p-values below 0.05, thus qualifying them for subsequent machine learning modeling.

Our comparative analysis through stratified 5-fold cross-validation revealed Logistic Regression as the most effective model. It excelled in recall (72%), precision (74%), and F1-score (71%), with an AUC of 80% that underscored its strong discriminative capacity. These findings (Table S2) comprehensively detail the model’s performance across various metrics, underscoring its robustness and reliability for predicting attendance at intensive schools based on behavioral data.

SHAP analysis

Our analysis utilized SHAP to understand the impacts of features on our predictive model’s decisions. The SHAP summary plot, (Fig. 3), highlights how lower values of viewing distance and shorter far viewing episodes increase the likelihood of intensive school attendance.

Figure 3.

SHAP summary plot depicting the impact of behavioral features on predictions of intensive school attendance. The left side of the figure (scatter plot) displays SHAP values for each feature. Positive SHAP values increase the likelihood of attendance. Red points represent higher feature values, and blue points represent lower feature values. Each dot represents a single instance (participant). For example, if a red dot is positioned on the right side of the graph, a higher value of that feature increases the model’s belief that the participant attended an intensive school educational system. The right side of the figure (bar chart) ranks features by their average impact, measured across all instances, with bar length indicating influence strength.

We investigated how feature values influence predictions about intensive school attendance and found that less than 1% long far-viewing episodes and near-viewing distances shorter than 44 cm increase the likelihood of intensive school attendance. This relationship was explored using SHAP dependence plots (Appendix 3, Figure A1), which illustrate the impact of feature values on model predictions.

Discussion

This pilot study explored the use of machine learning algorithms to analyze visual behaviors and assess their feasibility in predicting participants’ educational background. Findings confirm the association between specific near-viewing behaviors and ultra-Orthodox men, who are graduates of an intensive educational system. These results lay the groundwork for further studies that explore the potential causal links between specific behavioral patterns and their impact on the onset and progression of myopia. By identifying and quantifying specific behavioral patterns, such algorithms could eventually be used to monitor children during critical developmental periods when myopia typically onsets and progresses leading to insights regarding the role of visual behavior in the pathogenesis of myopia.

This pilot study focused on college students who had attended intensive (ultra-Orthodox) schools or standard (non-ultra-Orthodox) schools prior to college. Boys who attended intensive school systems begin their education as early as age three, attending schools that emphasize extensive reading of religious texts with small fonts and with a significantly higher number of study hours compared to standard schools²⁰. In contrast, students who attended standard schools start formal education at age six²² attending schools that follow a state curriculum with a broader range of subjects, including arts, mathematics, and science⁵⁰. The advantage of studying this population is its genetic homogeneity and geographical concentration, which minimizes confounding variables such as ethnicity and geographic location.

As hypothesized, those who attended intensive school systems exhibited significantly different near work behavior in three aspects: (1) the percentage of time spent at near viewing, (2) the average viewing distance, and (3) the patterns of continuous episodes of near viewing. Specifically, those who attended the intensive school system spent more time viewing things at very near distances and less time on intermediate near viewing, and their near viewing distance was significantly shorter than the group who attended a standard school system. Additionally, during near work, the graduates of the intensive school system took fewer long breaks and spent less time on them, while taking more frequent short breaks and spending more time on these compared to the graduates of the standard school system.

The machine learning model developed to predict whether an individual had attended an intensive school system based on objectively measured near viewing behaviors showed that the most influential features that increased the likelihood of classifying a participant’s attendance to an intensive school system were related to two aspects of near work behavior: shorter near-viewing distances and fewer long breaks during near work. The ability to predict the participant’s school system based on current near viewing behavior suggests that the distinct educational demands in kindergaren through high school (approximately ages 3–18 years) are associated with different near viewing behaviors that persisted into adulthood.

While the current study did not address causality between near-viewing and myopia, one may apply abductive reasoning to propose that the distinct near-viewing behaviors observed in graduates of intensive educational systems may contribute to differences in refractive status and the prevalence of myopia observed in this population level^{8,14,23,51,52}. However, future research needs to be carried out to effect alone is measured relative to other known correlations with a magnitude and compared against a baseline.

The distinct educational systems examined in this study, characterized by significantly different near viewing demands, may help explain trends seen in East Asia, where high educational demands and intensive near work^12,21 from an early age are linked to elevated rates of myopia⁵³. The parallels between the Israeli intensive educational system and those in countries such as China, South Korea, and Japan highlight the global relevance of these findings. By examining the shared educational practices and their impact on near-viewing behaviors, this study provides insights applicable beyond the specific populations studied here. However, in general, categorizing educational systems or populations into two broad groups may not fully capture the complexity of educational experiences and their impact on visual behavior. However, it should be noted that the current study did not assess causal relationship between near viewing behavior and myopia. Future research should investigate this potential link, considering that educational environments exist on a spectrum, with variations in curriculum structure, teaching methods, and extracurricular demands that could influence near viewing behaviors and myopia development.

This study examined near work behavior in college students, based on the assumption that near work habits formed in childhood may persist into adulthood. This study confirms near viewing patterns observed in a previous laboratory-based study in young adults³⁰. Findings from both studies reveal similar behavioral trends across these two groups, suggesting consistent patterns in both controlled and real-life conditions. In the previous laboratory-based study, participants who had attended the intensive school system read books and viewed tablets at closer distances than participants who had attended a standard school system³⁰. The average near work distances for different tasks ranged from 31.4 cm to 41.7 cm. In the current study, the average near-viewing distance was 45 cm. While we did not collect information on specific tasks performed at these distances, we found a significant difference of approximately 7 cm between students who had attended intensive school systems (41 cm) and those from standard school systems (48 cm).

Closer near viewing distances among students who had attended intensive school systems may be associated with early exposure to reading small text, which is typical in their religious books²⁰. Reading smaller font sizes may require closer viewing distances to increase the angular size of the text⁵⁴ a pattern also observed in Chinese children, where smaller font was associated with shorter viewing distances⁵⁵. These reading habits, established in childhood, may persist into adolescence and adulthood, contributing to the closer working distances seen in this group.

Research on educational demands has suggested that specific aspects of near work, in addition to near-viewing distance, such as the duration of near viewing and the frequency of breaks, may influence the development of myopia more than the total time spent on near work^18,19.

In the current study, the time spent viewing at near distances (10 and < 40 cm) and intermediate distances (40 and < 100 cm) was found to differ between students who had attended intensive versus standard school systems. Students who had attended intensive school systems tend to spend more time viewing at short distances, while students who had attended standard school systems spend more time viewing at intermediate distances. Further research should focus on understanding how the percentage of viewing time at each distance impacts myopia.

There were not significant differences in the overall percentage of breaks (i.e. far viewing breaks during near work) taken by the students who had attended either intensive school or standard school systems. However, when looking at the characteristics of the breaks, distinct behavioral differences emerged. Students who had attended intensive school tended to take more short break episodes and spend more time in short breaks, while those who had attended standard school tended to take longer breaks and spend more time in long breaks. Interestingly, Wallman and colleagues demonstrated that 2-minute interruptions of visual experience significantly reduced or prevented the development of myopia in chicks⁵⁶. This underscores the importance of far-viewing episodes in myopia progression. Also in our study we found that far viewing episodes longer than 5 min were associated with standard school systems. Further research should focus on understanding how viewing breaks impact myopia progression in humans.

Ultra-Orthodox graduates of an intensive schools system are known to have a higher prevalence and degree of myopia compared to non-ultra-Orthodox males who attended standard schools, which is often attributed to the intense near work demands in the intensive school educational system^23,52. The myopia prevalence among ultra-Orthodox participants who had attended intensive school educational systems in this study (82%) was not significantly different from those who had attended non-ultra-Orthodox systems (69%). However, there was a significantly higher degree of myopia among participants from intensive school educational systems. Additionally, the prevalence of high myopia was greater among ultra-Orthodox participants who had attended an intensive school educational system compared to those who had attended a standard school system. The observed myopia prevalence aligns with previous findings from our lab (87%) and reports in ultra-Orthodox Jewish males recruited into the army (82%)²³. The prevalence of myopia among participants from non-ultra-Orthodox educational backgrounds (69%) in this study closely matches rates observed in our previous study (68%)³⁰ and studies of male and female college students in Israel (66%)⁵⁷. However, it is higher than the myopia rate reported among Jewish males recruited into the army²³. It is important to note that college students, in general, tend to have higher prevalence rates of myopia compared to the general population⁵⁷. This may be due to the academic demands and near work associated with higher education, which can exacerbate myopia progression. This context highlights that the high myopia rates observed in both groups in this study may be influenced by the college student population itself. We speculate that myopia prevalence between the two groups studied here may have been influenced by the recruitment process, as all participants were college students. Importantly, the high prevalence of myopia among participants in both groups, combined with the similar myopia rates between the groups, enables a more balanced investigation of near-work behavior across educational backgrounds.

It is possible that the different viewing behaviours observed were a result of the ultra-Orthdox cohort having a significantly more myopic refractive error. However, no correlation was found between the degree of myopia and either the working distance or the duration of near viewing. This is similar to findings from previous studies in young adults²⁸ and children^29,55 which also showed no differences in near viewing distance between myopes and non-myopes for both in-lab tasks and for real world tasks⁴¹. However, another study using a near-work analyzer reported that high myopes had significantly shorter reading distances compared to low and non-myopes²⁵.

Machine learning analysis demonstrated that participants’ elementary and high school educational system could be predicted by their current adult viewing behavior, despite the relatively small sample size of 64 participants. Among the models tested, logistic regression performed best, achieving an AUC of 80%, precision of 74%, and recall of 72%. The ability of the model to achieve such high performance underscores the predictive strength of objective behavioral measurements, such as viewing distances and the percentage of long far-viewing episodes. Our focus on purely objective behavioral variables distinguishes this work from much of the existing literature^58–61 on educational outcomes, where survey-based or clinical measures are often used. The model’s ability to differentiate between participants from intensive school and standard school educational backgrounds, without reliance on non-behavioral features like refractive error or myopia, highlights the distinctiveness of visual behavior in these groups.

The SHAP analysis added further granularity by identifying specific cutoff values that distinguished the groups. For example, near-viewing distances shorter than 44 cm and far-viewing episodes longer than 5 min occurring less than 1% of the time were strongly associated with intensive school education.

Limitations of the study include the following: Participants were young adults with stable refractive status. Studying children in their habitual environments will provide more insight into behaviors during the critical period when myopia develops and progresses. Additionally, the study had only a small number of non-myopic participants, which could limit generalizability. However, the primary aim of this study was not to compare visual behavior between myopes and non-myopes; therefore, the limited number of non-myopic participants is less critical in this context. Future research exploring the relationship between myopia and near viewing behavior would benefit from a more balanced and diverse representation of refractive error groups to better assess behavioral differences in relation to myopia. Another limitation was the age difference between participants who attended intensive school educational systems and standard school systems, reflecting cultural differences in education timing; intensive school graduates tend to continue studying directly after high school, while standard school graduates typically attend university after military service. However, since myopia progression tends to stabilize after age 18⁶², this age difference is unlikely to have significantly affected the results. Cycloplegic refraction was not performed in this study, which may lead to an overestimation of myopia. However, in adults, this effect is expected to be minimal. Cycloplegia is particularly important in children to ensure accurate classification into refractive error categories⁶³. A recent study in young adult myopes found a difference between cycloplegic and noncycloplegic refraction, but the mean difference was relatively small, averaging − 0.2 D (− 0.25 to − 0.14 D)⁶⁴. Given that this study does not assess myopia progression or compare visual behavior between myopes and non-myopes, the use of cycloplegia was deemed less critical. Another consideration is the small sample size of the participants. However, as this is a pilot study, one of its key objectives was to evaluate the feasibility of using machine learning. Despite the limited sample, the machine learning model demonstrated strong performance, supporting its potential for future studies with larger populations using objective wearable sensors to collect behavioral data. The categorization of educational systems into two simplified groups—intensive and standard—does not reflect the full complexity and diversity of educational systems in Israel. These systems include distinctions such as ultra-Orthodox, religious, and secular systems, each with unique educational demands. Additionally, within the non-ultra-Orthodox category, there are both religious and secular subgroups, which have different rates of myopia despite sharing similar school systems. However, non-ultra-Orthodox men attending religious and secular schools experience no significant differences in the intensity of educational demands^20,22. Future research should aim to identify and include more nuanced features to provide a more comprehensive understanding of how specific educational environments influence near-viewing behaviors and myopia development.

Furthermore, although this study focused on educational practices, it is important to acknowledge that other lifestyle factors associated with strict ultra-Orthodox observance such as reduced outdoor exposure, specific dietary habits, and cultural norms¹⁰ may also contribute to differences in myopia prevalence. These potential confounding factors were not measured in the current study and should be considered in future investigations aiming to isolate the contribution of educational demands to myopia development.

The Clouclip device used in the study also presented limitations. It enters sleep mode if no movement is detected for 40 s, resulting in about 2.5 min of lost data per occasion. However, there were no significant differences between the groups in this respect. Another limitation is the potential misalignment between the clouclip’s beam and the participants’ actual line of sight, as eye position was not tracked. The device’s sampling rate of every 5 s may also miss short viewing breaks, but this limitation was consistent across both groups and unlikely to have influenced trends. Future experiments should be performed in other populations with different educational systems to assess whether the results observed here are consistent across other intensive education systems to provide a broader understanding of how educational environments impact near viewing behaviors and fundamental links between behavior and myopia development.

Conclusion

In conclusion, college students who had attended an intensive educational system exhibited distinct near viewing behaviors compared to those who had attended a standard educational system. Integration of machine learning enabled prediction of an individual’s educational system based solely on their viewing behavior. Such behaviors may be associated with the high prevalence and degree of myopia in this population. However, myopia causality cannot be established based on this study, and further research is needed to explore these relationships. While this study focuses on specific populations, the parallels between the intensive educational systems of the ultra-Orthodox population and those in certain Asian countries, as opposed to the Western-style systems of the non-ultra-Orthodox, suggest that these findings may have broader relevance and can ultimately be applied to inform strategies to mitigate myopia in similar educational contexts worldwide.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization, RD, ES, AGS, AG, LAO; Methodology, RD, ES, AGS, AG, LAO; Formal Analysis, RD, AG; Investigation, RD; Data Curation, RD, AG; Writing – Original Draft Preparation, RD, AG; Writing – Review & Editing, RD, ES, AGS, AG, LAO.

Funding

This study was supported by the United States-Israel Binational Science Foundation 2019053.

Data availability

The datasets used and/or analysed during the current study available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval

All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee (Jerusalem Multidisciplinary College Ethics Committee) and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.

Informed consent

Informed consent was obtained from all individual participants included in the study.