Abstract
Aims
Thyroid eye disease (TED) is an autoimmune orbital disorder that diminishes the quality of life (QOL) in affected individuals. Graves’ ophthalmopathy (GO)-QOL questionnaire effectively assesses TED’s effect on patients. This study aims to investigate the factors influencing visual functioning (QOL-VF) and physical appearance (QOL-AP) scores in Chinese TED patients using innovative data analysis methods.
Methods
This cross-sectional study included 211 TED patients whose initial visit to our clinic was from July 2022 to March 2023. Patients with previous ophthalmic surgery or concurrent severe diseases were excluded. GO-QOL questionnaires, detailed medical histories and clinical examinations were collected. The distribution of GO-QOL scores was analyzed, and linear regression and machine learning algorithms were utilized.
Results
The median QOL-VF and QOL-AP scores were 64.29 and 62.5, respectively. Multivariate linear regression analysis revealed age (P = 0.013), ocular motility pain (P = 0.012), vertical strabismus (P < 0.001) and diplopia scores as significant predictors for QOL-VF. For QOL-AP, gender (P = 0.013) and clinical activity (P = 0.086) were significant. The XGBoost model demonstrated superior performance, with an R2 of 0.872 and a root mean square error of 11.083. Shapley additive explanations (SHAP) analysis highlighted the importance of vertical strabismus, diplopia score and age in influencing QOL-VF and age, clinical activity and sex in QOL-AP.
Conclusion
TED significantly affects patient QOL. The study highlights the efficacy of XGBoost and SHAP analyses in identifying key factors influencing the QOL in TED patients. Identifying effective interventions and considering specific demographic characteristics are essential to improving the QOL of patients with TED.
Introduction
Thyroid eye disease (TED) is an autoimmune orbital disorder and the most important extrathyroidal manifestation of Graves’ disease (GD) (1). Approximately 50% of patients with GD develop TED, and these patients often suffer from eyelid redness and swelling, pain, proptosis, ocular motility restriction, diplopia and, in some cases, visual deterioration. These persistent signs and symptoms seriously affect the quality of life (QOL).
A person’s QOL reflects their well-being and ability to function in life, influenced by factors such as vision loss, financial strain and social relationships (2). Assessing TED merely through physical examination often underestimates its effect (3). To better evaluate the QOL in TED patients, a Graves’ ophthalmopathy (GO)-QOL questionnaire, developed by Terwee and coworkers (4), is commonly used. This questionnaire consists of 16 questions focusing on two specific aspects: eight items assessing visual functioning (QOL-VF) and eight items addressing physical appearance (QOL-AP). Understanding the effect of TED on patient well-being is crucial for identifying factors that affect QOL, thus enabling the development of new strategies to improve disease management (5).
The QOL in TED patients can be influenced by many factors, which can offer valuable insight for treating the disease. However, results from earlier studies relying on the GO-QOL questionnaire have varied because of factors such as small sample sizes, data quality issues and differences in analytical methods (6, 7, 8). Consensus regarding the relative importance of factors such as sex, age, activity level and disease severity has yet to be achieved. To better understand the factors affecting GO-QOL scores, more comprehensive clinical data are needed.
Machine learning (ML) is a sophisticated computational method used for analyzing data, handling missing values and tackling overfitting and underfitting issues (9, 10). ML, often accompanied by Shapley additive explanations (SHAP), a method for identifying the characteristics of indicators in prediction models, has been employed in the analysis of factors influencing the QOL (11, 12). Previous studies using GO-QOL have predominantly relied on univariate and multivariate analyses. ML emerges as an underutilized yet promising method for predicting GO-QOL scores in patients with TED. Therefore, we embarked on implementing ML algorithms for regression modeling and SHAP for feature analysis of the QOL in TED patients.
The current study aimed to better understand the factors influencing the QOL in patients with TED, using a Chinese population as an example. A questionnaire survey was provided to TED patients visiting our clinics. Detailed clinical data of the patients were collected and studied, employing various ML models to identify and interpret the factors influencing the QOL.
Methods
Patients
This retrospective study was approved by the ethics committee (see the ‘Ethics approval consent to participate’ section for details). TED patients visiting the outpatient clinic at our department of ophthalmology for the first time from July 2022 to March 2023 were recruited for this study. The inclusion criteria were as follows: (i) TED diagnosis according to the European Group on Graves’ Orbitopathy (EUGOGO) clinical guidelines at our outpatient clinic (13); (ii) aged between 18 and 75 years; and (iii) completion of the GO-QOL self-administered questionnaires at the outpatient clinic before undergoing examination. The exclusion criteria were as follows: (i) previous orbital decompression surgery or strabismus surgery; (ii) presence of systemic comorbidities such as diabetes, hepatitis or other ocular diseases such as cataracts and glaucoma, etc.; and (iii) lack of key medical history or clinical examination records. In total, 67 subjects were excluded (17, 29 and 21 for the three exclusion criteria).
GO-QOL questionnaires
The QOL in TED patients was assessed using the GO-QOL questionnaire, which was routinely administered before appointments. A small number of patients (less than 30 individuals) declined to complete the questionnaire because of time constraints or other reasons. The questionnaire, sourced from the EUGOGO website, was translated into Chinese and comprised two subscales: QOL-VF and QOL-AP, with each item rated on a three-point scale (1 = severely limited, 2 = a little limited and 3 = not limited at all). Researchers conducted face-to-face interactions to explain, administer and record the questionnaire. Scores ranged from 8 to 24, which were then converted to a scale of 0–100. In this conversion, 0 represented the worst QOL and 100 represented the best. The formula used for conversion was as follows: (total scores − X)/(2 × X) × 100, where X meant the number of completed items (4). QOL-VF and QOL-AP scores for each patient were obtained as response variables.
Demographic and clinical profiles
To assess the factors affecting patient QOL, we conducted a retrospective analysis, gathering demographic and clinical data from medical records, including sex, age, education and disease duration since the first TED-related ocular signs or symptoms. Treatment histories such as intravenous steroids and other interventions (radiotherapy, local injection, oral steroids and immunosuppressants) were noted. We also evaluated TED activity using the clinical activity score (CAS) and orbital MRI results. Patients with a CAS ≥ 3 or noticeable inflammatory changes on orbital MRI were categorized as having active TED, while those without these criteria were classified as inactive TED (14). The severity was determined using the EUGOGO classification system, dividing patients into mild, moderate-to-severe and sight-threatening categories.
Patient clinical profiles, such as best-corrected visual acuity, palpebral fissure height and proptosis as measured by Hertel exophthalmometry, were collected by a single ophthalmologist (15). The thyroid status was considered stable if serum-free thyroxine (FT4), serum free triiodothyronine (FT3) and thyroid stimulating hormone levels were all within the normal range. Recorded periocular inflammatory signs included five objective signs of CAS: eyelid redness, eyelid swelling, conjunctival redness, conjunctival swelling and caruncle or plica swelling. Ocular motility restriction was considered positive if the patient experienced difficulty in moving the eye in any direction. Strabismus was measured by prism cover testing in different gaze directions. Those with a horizontal deviation greater than 15 PD or a vertical deviation greater than 10 PD were considered to have strabismus (16). The Gorman subjective diplopia score was also evaluated, which ranges from 0 to 3: no diplopia (0), intermittent diplopia in primary gaze when awakening or tired (1), diplopia at extremes of gaze (2) and continuous diplopia in primary or reading gaze (3). Additionally, patients were asked about symptoms, including spontaneous retrobulbar pain, ocular motility pain, photophobia, excessive tearing and dry eyes, with responses recorded as yes or no.
Predictive modeling with machine learning algorithms
We individually implemented various machine learning models, including ensemble methods such as XGBoost (eXtreme Gradient Boosting), AdaBoost (adaptive boosting), random forest and gradient boosting; decision-based algorithms such as decision tree; and regularization techniques such as ridge, lasso regression and elastic net (17, 18, 19, 20). Additionally, we utilized support vector regression, K-nearest neighbors (K-neighbors) and multilayer perceptron regression (MLP regression) to address different predictive challenges (21). According to an 8:2 ratio, the 211 patients were randomly divided into a training group and a verification group, consisting of 169 and 42 patients, respectively.
To analyze the contributions of each feature in the optimal prediction model, we utilized the SHAP methodology through the ‘SHAP’ Python library. SHAP values were computed to illustrate how each variable influences the model’s output (22). By using the ‘SHAP. Explainer’ function, we fitted the model and derived SHAP values, which gauge the expected change in the model output per feature. These values facilitated the creation of a feature importance plot, demonstrating the magnitude and direction of each feature’s effect on the predictions. This process helped identify key features that significantly affect outcomes. For a more detailed explanation of these machine learning algorithms in predictive modeling, see the Supplementary Method (see the section on Supplementary materials given at the end of the article).
Statistical analysis
The baseline characteristics of TED patients were outlined using descriptive statistics and frequency tables. Indicators were categorized based on the variables under study. Categorical variables were summarized using frequency tables. Numerical variables following a normal distribution were presented as mean ± standard deviation (SD), while those deviating from normality were described using the interquartile range and median. Spearman rank correlation or Pearson correlation analysis was employed to assess the correlation between numerical indicators and the GO-QOL score, while the Mann–Whitney test was utilized for categorical indicators. For potential influencing factors of the GO-QOL score identified in the univariate analysis, multivariate linear regression was conducted to evaluate the effect of each influencing factor. Statistical significance was set at P < 0.05. The analyses were conducted using SPSS 19.0.
To assess the predictive capabilities of our models, we utilized several statistical metrics, including the mean absolute error, mean square error (MSE), R2, variance and root MSE (RMSE). This facilitated a comprehensive comparison between the ML algorithms and the conventional statistical approach of linear regression.
Results
Characteristics of participants
The demographic and clinical characteristics of the study participants are shown in Table 1. Among the participants, 138 cases (65.4%) were female, with a mean age at study entry of 43.6 ± 13.2 years. The median duration of TED was 12 months, ranging from 6 to 26 months. Approximately equal proportions of TED patients exhibited stable (51.2%) and unstable (48.8%) thyroid activity statuses. The majority of patients had not previously received intravenous steroids (151 cases, 71.6%) or other TED-related treatments (165 cases, 78.2%, including oral steroids, local steroid injections, immunosuppressants and radiotherapy). Within the cohort, 100 patients (47.4%) had active TED. Most patients (144 cases, 68.3%) were diagnosed with moderate-to-severe TED. The most frequently reported periocular inflammatory sign was eyelid swelling (136 cases, 64.5%). Common symptoms included photophobia (60.2%) and excess tearing (58.8%). Additionally, 58.8% of patients presented with varying degrees of diplopia, with constant diplopia accounting for 37% of cases, representing the most prevalent category.
Demographic and clinical characteristics of patients with TED. Data are presented as n (%) or mean ± SD or median (Q1, Q3).
| Characteristics | Values |
|---|---|
| Sex | |
| Male | 73 (34.6) |
| Female | 138 (65.4) |
| Age (years) | 43.6 ± 13.2 |
| Education (years) | 15 (9,16) |
| Disease duration (months) | 12 (6, 26) |
| Thyroid status | |
| Stable | 108 (51.2) |
| Unstable | 103 (48.8) |
| TED treatment history | |
| Intravenous steroids | 60 (28.4) |
| Other treatment | 46 (21.8) |
| Activity | |
| Inactive | 111 (52.6) |
| Active | 100 (47.4) |
| Severity | |
| Mild | 25 (11.8) |
| Moderate-to-severe | 144 (68.3) |
| Sight-threatening | 42 (19.9) |
| Best-corrected visual acuity | 1.0 (0.8,1.0) |
| Proptosis (mm) | 19.50 (18.00, 21.25) |
| Palpebral fissure height (mm) | 10.0 (8.5,11.0) |
| Periocular inflammatory signs | |
| Eyelid redness | 14 (6.6) |
| Eyelid swelling | 136 (64.5) |
| Conjunctival redness | 61 (28.9) |
| Conjunctival swelling | 59 (28.0) |
| Caruncle or plica swelling | 71 (33.6) |
| Discomfort symptoms | |
| Spontaneous retrobulbar pain | 57 (27.0) |
| Ocular motility pain | 37 (17.5) |
| Photophobia | 127 (60.2) |
| Excess tearing eyes | 124 (58.8) |
| Dry eyes | 89 (42.2) |
| Ocular motility restriction | 148 (70.1) |
| Strabismus | |
| Vertical strabismus | 58 (27.5) |
| Horizontal strabismus | 66 (31.3) |
| Diplopia score | |
| 0: no diplopia | 87 (41.2) |
| 1: intermittent diplopia | 8 (3.8) |
| 2: gaze-dependent diplopia | 38 (18.0) |
| 3: constant diplopia | 78 (37) |
TED, thyroid eye disease.
QOL impairment in TED patients
Patient QOL was affected to varying extents by TED (Fig. 1A). The median QOL-VF score stood at 64.29 (42.86, 93.75), with 41.23% (87 patients) achieving scores above 75, indicating relatively good QOL-VF. Regarding QOL-AP, the median score was 62.5 (43.75, 87.50), with 36.02% (76 patients) scoring above 75 and 13.27% (28 patients) falling below 25. Additionally, 11 patients scored below 25 in both categories, while 35 patients scored above 75 in both. In Fig. 1B, inactive patients dominate the highest QOL-VF score category (75, 100). Conversely, Fig. 1C reveals a higher proportion of active patients in the uppermost QOL-AP score range. In Fig. 1D and E, a greater percentage of patients with mild conditions fall within the (75–100) range for both VF and AP, whereas those with sight-threatening disease are most frequently found in the (0–25) score range. In summary, most patients experienced mild to moderate impairment in their QOL scores.
Distribution of QOL-VF and QOL-AP among patients with TED. (A) Distribution of QOL-VF and QOL-AP scores. Score ranges are divided into (0–25), (25–50), (50–75) and (75–100). Distribution of QOL-VF (B) and QOL-AP (C) scores in relation to activity levels (active vs inactive). Distribution of QOL-VF (D) and QOL-AP (E) scores across disease severities (mild, moderate-to-severe and sight-threatening). QOL, quality of life; QOL-VF, visual functioning; QOL-AP, appearance.
Citation: European Thyroid Journal 14, 1; 10.1530/ETJ-24-0292
Results of the univariate and multivariate analysis
Visual functioning subscale
Univariate analyses disclosed that 17 out of 26 variables were significantly associated with QOL-VF (Table 2). Female patients exhibited higher QOL-VF scores (66.76 ± 31.65) compared with their male counterparts (52.16 ± 33.65, P = 0.002). QOL-VF scores showed a negative correlation with age (P < 0.001) and a positive correlation with years of education (P = 0.008) and disease duration (P = 0.002). Patients who previously underwent intravenous steroids (P < 0.001) or other treatments (P = 0.007) had significantly lower QOL-VF scores.
Results of univariate analysis.
| QOL-VF | QOL-AP | |||||
|---|---|---|---|---|---|---|
| Mean ± SD | t | P | Mean ± SD | t | P | |
| Sex | 0.002* | 2.77 | 0.006* | |||
| Male | 52.16 ± 33.65 | −3.11 | 69.18 ± 27.15 | |||
| Female | 66.76 ± 31.65 | 57.74 ± 29.25 | ||||
| Age (years) | −6.76 | <0.001* | 0.128 | 1.5 | 0.134 | |
| Education (years) | 2.69 | 0.008* | 0.002 | −0.07 | 0.945 | |
| Disease duration (months) | 3.18 | 0.002* | −0.097 | 0.82 | 0.412 | |
| Thyroid status | 1.46 | 0.145 | 0.05 | 0.957 | ||
| Stable | 64.96 ± 32.91 | 61.81 ± 29.85 | ||||
| Unstable | 58.31 ± 33.01 | 61.59 ± 28.22 | ||||
| Intravenous steroids | −4.87 | <0.001* | −1.8 | 0.074 | ||
| Yes | 45.00 ± 31.01 | 56.04 ± 35.73 | ||||
| No | 68.35 ± 31.54 | 63.95 ± 25.63 | ||||
| Other treatment | −2.74 | 0.007* | −1.37 | 0.171 | ||
| Yes | 50.08 ± 35.55 | 56.52 ± 36.42 | ||||
| No | 64.96 ± 31.68 | 63.14 ± 26.51 | ||||
| Activity | −4.53 | <0.001* | 2.09 | 0.038* | ||
| Inactive | 71.06 ± 28.57 | 57.77 ± 28.58 | ||||
| Active | 51.33 ± 34.69 | 66.06 ± 28.97 | ||||
| Severity | −0.07 | 0.313 | ||||
| Mild | 75.77 ± 31.87 | 65.75 ± 28.02 | ||||
| Moderate–severe | 66.91 ± 30.19 | −1.35 | 0.179 | 62.07 ± 28.69 | −0.59 | 0.559 |
| Sight-threatening | 35.52 ± 30.04 | −5.25 | <0.001* | 58.04 ± 30.84 | −1.05 | 0.294 |
| Best-corrected visual acuity | 6 | <0.001* | 0.132 | 1.4 | 0.163 | |
| Proptosis (mm) | 1.08 | 0.282 | −0.063 | −0.17 | 0.864 | |
| Palpebral fissure height (mm) | 1.93 | 0.055 | 0.028 | 0.77 | 0.441 | |
| Eyelid redness | 0 | 0.998 | 0.88 | 0.379 | ||
| Yes | 61.69 ± 31.38 | 68.30 ± 29.86 | ||||
| No | 61.71 ± 33.24 | 61.23 ± 28.95 | ||||
| Eyelid swelling | −2.45 | 0.015* | 0.01 | 0.99 | ||
| Yes | 57.62 ± 32.40 | 61.72 ± 29.82 | ||||
| No | 69.13 ± 33.14 | 61.67 ± 27 .63 | ||||
| Conjunctival redness | −1.47 | 0.144 | 0.22 | 0.824 | ||
| Yes | 56.49 ± 32.80 | 62.40 ± 30.16 | ||||
| No | 63.84 ± 33.02 | 61.42 ± 28.61 | ||||
| Conjunctival swelling | −2.16 | 0.032* | 0.02 | 0.133 | ||
| Yes | 53.90 ± 32.73 | 61.76 ± 32.85 | ||||
| No | 64.75 ± 32.78 | 61.68 ± 27.47 | ||||
| Caruncle or plica swelling | −1.82 | 0.071 | −1.51 | 0.184 | ||
| Yes | 55.94 ± 32.15 | 57.48 ± 30.45 | ||||
| No | 64.64 ± 33.23 | 63.84 ± 28.10 | ||||
| Spontaneous retrobulbar pain | −2.46 | 0.015* | −0.49 | 0.624 | ||
| Yes | 52.61 ± 33.56 | 60.09 ± 31.78 | ||||
| No | 65.08 ± 32.32 | 62.30 ± 27.98 | ||||
| Ocular motility pain | −3.47 | <0.001* | −1.03 | 0.307 | ||
| Yes | 45.01 ± 30.16 | 57.26 ± 32.56 | ||||
| No | 65.26 ± 32.62 | 62.64 ± 28.19 | ||||
| Photophobia | −1.88 | 0.061 | −1.02 | 0.307 | ||
| Yes | 58.25 ± 33.91 | 60.04 ± 29.41 | ||||
| No | 66.95 ± 31.18 | 64.21 ± 28.34 | ||||
| Excess tearing | −0.99 | 0.325 | 0.3 | 0.767 | ||
| Yes | 59.83 ± 34.05 | 62.20 ± 28.89 | ||||
| No | 64.39 ± 31.57 | 60.99 ± 29.30 | ||||
| Dry eyes | −1.64 | 0.103 | 0.79 | 0.429 | ||
| Yes | 57.37 ± 33.55 | 63.55 ± 27.57 | ||||
| No | 64.88 ± 32.45 | 60.35 ± 30.03 | ||||
| Ocular motility restriction | −3.99 | <0.001* | −0.33 | 0.745 | ||
| Yes | 55.98 ± 32.82 | 61.28 ± 30.42 | ||||
| No | 75.17 ± 29.72 | 62.70 ± 25.55 | ||||
| Vertical strabismus | −9.2 | <0.001* | −0.05 | 0.958 | ||
| Yes | 32.96 ± 28.87 | 61.53 ± 30.25 | ||||
| No | 72.61 ± 27.59 | 61.76 ± 28.60 | ||||
| Horizontal strabismus | −4.27 | <0.001* | 0.69 | 0.494 | ||
| Yes | 47.86 ± 32.89 | 63.73 ± 26.16 | ||||
| No | 68.02 ± 31.25 | 60.78 ± 30.24 | ||||
| Diplopia score | −0.47 | 0.022 | 0.752 | |||
| 0: no diplopia | 80.07 ± 26.01 | 61.14 ± 27.93 | ||||
| 1: intermittent diplopia | 35.94 ± 29.30 | −4.11 | <0.001* | 64.83 ± 32.03 | 0.34 | 0.731 |
| 2: gaze-dependent diplopia | 57.01 ± 31.72 | −4.08 | <0.001* | 61.68 ± 30.18 | 0.1 | 0.924 |
| 3: constant diplopia | 46.17 ± 30.93 | −7.48 | <0.001* | 62.02 ± 30.18 | 0.19 | 0.846 |
QOL, quality of life; QOL-AP, appearance; QOL-VF, visual functioning.
Statistical significance (P < 0.05).
Patients with inactive disease had a higher mean QOL-VF score (71.06 ± 28.57) compared with those with active disease (51.33 ± 34.69, P < 0.001). A decreasing trend in QOL-VF subscale scores was observed as TED severity progressed from mild to sight-threatening TED (P < 0.001). These QOL-VF scores correlated with eyelid (P = 0.015) and conjunctival swelling (P = 0.032). Objective ocular signs such as best-corrected visual acuity, ocular motility restriction, vertical strabismus, horizontal strabismus and diplopia score were significantly correlated with QOL-VF scores (P < 0.001). Additionally, spontaneous retrobulbar pain (P = 0.015) and ocular motility pain (P < 0.001) were also found to be significantly correlated with QOL-VF scores.
Multivariate analysis of QOL-VF scores was conducted using a linear regression model (R2 = 0.4729). The results revealed a significant predictive value for age (P = 0.013), ocular motility pain (P = 0.012), vertical strabismus (P < 0.001), diplopia scores, including intermittent diplopia (P < 0.001), gaze-dependent diplopia (P = 0.029) and constant diplopia (P = 0.024) (Table 3). According to the definitions suggested by Bradley and coworkers, there was a significant ceiling effect in the assessment of QOL-VF (40/211 = 18.96% > 15%) (23).
Multivariate analysis of the predictors of quality of life-visual functioning†.
| Regression coefficient | Standard error | t | P | |
|---|---|---|---|---|
| Male | −0.17 | 3.78 | −0.05 | 0.963 |
| Age | −0.39 | 0.16 | −2.51 | 0.013* |
| Education (years) | −0.06 | 0.44 | −0.14 | 0.889 |
| Disease duration (months) | 0.07 | 0.04 | 1.8 | 0.073 |
| Activity | −1.02 | 4.31 | −0.24 | 0.813 |
| Severity | ||||
| Mild | ||||
| Moderate–severe | −7.44 | 5.68 | −1.31 | 0.191 |
| Sight-threatening | −11.82 | 9.49 | −1.25 | 0.215 |
| Intravenous steroids | −5.99 | 4.23 | −1.41 | 0.159 |
| Other treatment | 1.36 | 4.35 | 0.31 | 0.756 |
| Best-corrected visual acuity | 27.24 | 14.07 | 1.94 | 0.054 |
| Eyelid swelling | 3.81 | 4.09 | 0.93 | 0.353 |
| Conjunctiva swelling | 1.07 | 4.38 | 0.24 | 0.807 |
| Spontaneous retrobulbar pain | −4.39 | 4.34 | −1.01 | 0.313 |
| Ocular motility pain | −13.06 | 5.17 | −2.53 | 0.012* |
| Ocular motility restriction | −0.26 | 4.2 | −0.06 | 0.95 |
| Vertical strabismus | −24.58 | 4.63 | −5.31 | <0.001* |
| Horizontal strabismus | −0.48 | 4.09 | −0.12 | 0.907 |
| Diplopia score | ||||
| 0: no diplopia | ||||
| 1: intermittent diplopia | −33.69 | 9.61 | −3.5 | <0.001* |
| 2: gaze-dependent diplopia | −11.07 | 5.03 | −2.2 | 0.029* |
| 3: constant diplopia | −10.66 | 4.7 | −2.27 | 0.024* |
Statistical significance (P < 0.05).
Multiple linear regression model: n = 211, R2 = 0.4729.
Appearance subscale
QOL-AP scores were found to correlate with sex (P = 0.006) and disease activity (P = 0.038) (Table 2). Male patients achieved a significantly higher QOL-AP score (69.18 ± 27.15) compared with females (57.74 ± 29.25). Patients with active disease exhibited higher mean QOL-AP scores (66.06 ± 2.90) compared with inactive patients (57.77 ± 2.71). Although no significant correlation was observed between QOL-AP scores and disease severity (P = 0.313), a decreasing trend was noted as TED severity increased from mild to sight-threatening TED.
Multivariate analysis of QOL-AP scores was conducted using a linear regression model. The results revealed that sex, but not disease activity, represented the variable with predictive value for QOL-AP (P = 0.013) (Table 4). No ceiling effect was detected in the assessment of QOL-AP scores.
Evaluating the efficacy of predictive models built with machine learning algorithms
The performance metrics varied across the models, with R2 values ranging from −1.17 to 0.872 and RMSE values from 11.083 to 44.552, as detailed in Supplementary Tables 1 and 2. XGBoost emerged as the optimal model, showcasing leading performance metrics (R2 = 0.872, RMSE = 11.083) (Supplementary Tables 1 and 2). Significant variability was detected in the effectiveness of other models. For instance, multiple linear regression and AdaBoost demonstrated moderate performance, whereas decision tree and MLP regression were less effective.
Important features predicting GO-QOL investigated by SHAP
The beeswarm plots generated from the SHAP analysis, illustrating the ranking of feature importance for all 26 features within the XGBoost model, are depicted in Figs 2 and 3. These plots highlight the direction of impact and provide patient-specific information, including outliers.
SHAP summary plot of feature importance in predicting QOL-VF. Each point represents a patient, with colors indicating feature values (red = high value and blue = low value). The x-axis shows SHAP values, where positive values (right) indicate features increasing QOL-AP scores and negative values (left) indicate features decreasing QOL-AP scores. Features are ordered by their absolute effect on model predictions, with the most influential features at the top. SHAP, Shapley additive explanations.
Citation: European Thyroid Journal 14, 1; 10.1530/ETJ-24-0292
SHAP summary plot of feature importance to predict QOL-AP. Each point represents a patient, with colors indicating feature values (red = high value and blue = low value). The x-axis shows SHAP values, where positive values (right) indicate features increasing QOL-AP scores and negative values (left) indicate features decreasing QOL-AP scores. Features are ordered by their absolute effect on model predictions, with the most influential features at the top. SHAP, Shapley additive explanations.
Citation: European Thyroid Journal 14, 1; 10.1530/ETJ-24-0292
Visual functioning subscale
Vertical strabismus, diplopia score, age, best-corrected visual acuity, palpebral fissure height and ocular motility pain significantly affected the visual functioning subscale model. The SHAP analysis revealed that lower values of vertical strabismus, diplopia score and age were associated with higher positive SHAP values, indicating a positive influence on the predicted QOL-VF scores. Conversely, higher values of these features were associated with larger negative SHAP values, indicating a negative effect on the predicted QOL-VF scores (see Fig. 2). Specifically, the absence of vertical strabismus, lower diplopia scores or younger age was associated with improved vision-related QOL as predicted by the model.
Appearance subscale
Age, disease activity, sex, disease duration, best-corrected visual acuity and proptosis had the greatest effects on the appearance subscale model. Among these variables, disease activity and sex demonstrated a trend in which higher values corresponded to higher SHAP values, indicating increased predicted QOL-AP scores (refer to Fig. 3). Patients with active disease and male patients both showed an improvement in appearance-related QOL.
Discussion
This study detected varying degrees of QOL reduction among the investigated population of TED patients, a finding consistent with earlier studies. QOL-VF and QOL-AP scores were lower in TED patients compared with normative values. The factors influencing the QOL in TED patients included those deemed nonmodifiable, such as sex, age and disease duration, as well as modifiable factors, including best-corrected visual acuity, ocular motility pain, vertical strabismus and diplopia.
The study revealed that patients in the active phase experienced worse QOL-VF, while those in the inactive phase were more likely to have worse QOL-AP. These findings might result from persistent appearance changes in longer-duration TED, leading to psychological stress and social challenges. This finding suggests that patients with TED, as a chronic disease, require follow-up and evaluation beyond the active period (24). We observed a statistically significant decline in QOL-VF in sight-threatening disease but detected a progressive decline in the three patient groups categorized by severity. This trend was consistent with those reported by Zeng et al. (25). Overall, disease severity remains a significant factor in determining QOL-VF scores in TED patients.
Sex and age significantly affect both QOL-VF and QOL-AP. Women typically score lower on QOL-AP, while men are more affected in terms of QOL-VF. Societal expectations often place more emphasis on women’s appearance, increasing their concern about their looks (26). Conversely, men might experience lower QOL-VF because of more severe ocular symptoms and a worse prognosis (27). Additionally, QOL-VF tends to decrease with age, highlighting the importance of age-appropriate treatment options to improve visual function and self-care, particularly for the elderly.
SHAP analysis demonstrated considerable agreement (i.e., a negative correlation with QOL-VF) with multivariate linear regression results, underscoring the profound effect of vertical strabismus and diplopia on QOL-VF. Diplopia could make daily activities such as reading and driving challenging and increase the risk of injuries such as falls, especially in elderly patients (28). Similarly, the study by Ponto et al. (29) demonstrated a significant association between diplopia and reduced QOL-VF, which aligns with our findings. McBain et al. (30) assessed the QOL among 220 patients with strabismus, finding scores to be markedly lower than those of the healthy cohort. Our study further indicates a more pronounced effect from vertical strabismus. This stronger association may be because of the smaller vertical fusion range and potentially greater effect on stereopsis and depth perception. Vertical deviations might also be more noticeable in social interactions. However, these hypotheses require further research. All in all, our findings suggest the potential value of strabismus correction surgery in stable TED patients for improving their QOL.
Subjective discomfort symptoms reported by patients, particularly spontaneous retrobulbar pain and ocular motility pain, correlated with a reduced QOL-VF. This correlation aligns with earlier studies, which demonstrated that an increase in ocular pain significantly diminishes the QOL (2). It underscores the importance of acknowledging the disparities that exist between patients’ and physicians’ perspectives (31) and emphasizes the potential negative effect these disparities may have on therapeutic outcomes and doctor–patient communication.
Similar to the findings of Wang et al. (32), patients who never received intravenous steroids or other treatments had better QOL-VF, seemingly contradicting the expected benefits of these therapies. This unexpected outcome may be attributed to several factors: patients receiving treatment likely have more severe conditions; treatment may increase patients’ psychological stress and expectations; and the complex interplay between variables could explain why treatment history was not a significant predictor in multiple linear regression analysis. These findings underscore the need for a nuanced understanding of the effect of TED treatments on patient-reported outcomes.
Earlier studies largely utilized univariate and multivariate linear regression statistical analyses (33, 34). To our knowledge, this study validates, for the first time, the factors affecting the QOL in TED patients, utilizing machine learning. Accurately interpreting ML prediction models and intuitively presenting the results to clinicians have always been a challenge. Therefore, we applied SHAP values to XGBoost to achieve optimal predictive performance and interpretability. SHAP values assess the importance of all feature combinations on the output, providing consistent and locally accurate attribution values for each feature, enabling clinicians to more intuitively understand the decision-making process of this black-box model. In addition, this model presents several unique advantages over traditional linear models. For instance, in multivariate linear regression analysis, disease duration and best-corrected visual acuity were not significantly correlated with QOL-VF; however, the results of machine learning analysis demonstrated their influence on QOL-VF. This result suggests that linear regression methods failed to capture nonlinear relationships (35).
Several limitations affect our study. First, employing a cross-sectional design identifies factors affecting a particular TED patient’s QOL at only one time point, failing to establish symptom causality. Follow-up and prospective research remain essential for evaluating the longitudinal validity of the GO-QOL. Second, the current study sampled exclusively at a single institution with a referral patient population; thus, these results may not accurately represent the broader patient population. Larger, multicenter studies will be required for validation. Finally, our research can accommodate additional parameters as potential factors influencing the QOL in TED, such as tobacco smoking habits.
Our findings highlight the potential of ML approaches in analyzing GO-QOL, opening new avenues for future research. Longitudinal studies tracking the QOL before and after treatment could provide valuable insights into the efficacy of various interventions, particularly for chronic TED patients. Furthermore, this approach could be instrumental in evaluating emerging therapies, helping to tailor treatments to individual patient needs and potentially improving overall patient outcomes. Future research should also focus on validating these findings in larger, diverse cohorts and exploring the long-term implications of QOL changes in TED management.
Conclusion
The effect of TED on the QOL is highly variable, highlighting the need for physicians to better understand and tailor treatments based on patient-specific factors such as sex, age and disease duration. The effect of specific manifestations such as diplopia, visual acuity, pain and mobility restriction must be explored in greater detail. This study demonstrates the effectiveness of machine learning, especially XGBoost and SHAP, in analyzing QOL influences in TED.
Supplementary materials
This is linked to the online version of the paper at https://doi.org/10.1530/ETJ-24-0292.
Declaration of interest
The authors declare that there is no conflict of interest that could be perceived as prejudicing the impartiality of the work reported.
Funding
This work was supported by the National Natural Science Foundation of China (82388101, 81930024 and 82271122), the Science and Technology Commission of Shanghai (20DZ2270800), Shanghai Key Clinical Specialty, Shanghai Eye Disease Research Center (2022ZZ01003), and the Project of Hospital Management (YGA202303) from Shanghai Ninth People’s Hospital, Shanghai Jiao Tong University School of Medicine.
Author contribution statement
H Zhang and S Wu conceived and designed the study, performed data analysis and interpretation and wrote the manuscript. L Yang, C Fan, H Chen, H Wang, T Zhu, Y Li, J Sun and X Song contributed to data acquisition, analysis and interpretation. H Zhou, T J Smith and X Fan provided guidance throughout the study, supervising the research and critically revising the manuscript for important intellectual content.
Data Availability
Data sharing is not applicable to this article as no datasets were generated or analyzed during the current study.
Ethics approval consent to participate
This retrospective study was approved by the Ethics Committee of Shanghai Ninth People’s Hospital. Given the retrospective nature of this study, the ethics committee approved the waiver of individual informed consent.
Acknowledgements
We express our heartfelt gratitude to all the patients who participated in this study and the dedicated healthcare professionals who made this research possible.
References
- 1↑
Bartalena L & Tanda ML. Current concepts regarding Graves’ orbitopathy. J Intern Med 2022 292 692–716. (https://doi.org/10.1111/joim.13524)
- 2↑
Kahaly GJ, Petrak F, Hardt J, et al. Psychosocial morbidity of Graves’ orbitopathy. Clin Endocrinol 2005 63 395–402. (https://doi.org/10.1111/j.1365-2265.2005.02352.x)
- 3↑
Terwee CB, Dekker FW, Bonsel GJ, et al. Facial disfigurement: is it in the eye of the beholder? A study in patients with Graves’ ophthalmopathy. Clin Endocrinol 2003 58 192–198. (https://doi.org/10.1046/j.1365-2265.2003.01695.x)
- 4↑
Terwee CB, Gerding MN, Dekker FW, et al. Development of a disease specific quality of life questionnaire for patients with Graves’ ophthalmopathy: the GO-QOL. Br J Ophthalmol 1998 82 773–779. (https://doi.org/10.1136/bjo.82.7.773)
- 5↑
Smith TJ, Cockerham K, Barretto N, et al. Bridging and validation of the specific Graves’ ophthalmopathy quality of life (GO-QOL) questionnaire with health state utility values. Endocr Pract 2024 30 470–475. (https://doi.org/10.1016/j.eprac.2024.02.002)
- 6↑
Cockerham KP, Padnick-Silver L, Stuertz N, et al. Quality of life in patients with chronic thyroid eye disease in the United States. Ophthalmol Ther 2021 10 975–987. (https://doi.org/10.1007/s40123-021-00385-8)
- 7↑
Lee H, Roh HS, Yoon JS, et al. Assessment of quality of life and depression in Korean patients with Graves’ ophthalmopathy. Korean J Ophthalmol 2010 24 65. (https://doi.org/10.3341/kjo.2010.24.2.65)
- 8↑
Zhong SS, Liu ZY, Wu XS, et al. Influence factors for decreased quality of life in thyroid-associated ophthalmopathy patients. Guoji Yanke Zazhi 2017 17 311–316. (https://doi.org/10.3980/j.issn.1672-5123.2017.2.30)
- 9↑
Graham SA, Lee EE, Jeste DV, et al. Artificial intelligence approaches to predicting and detecting cognitive decline in older adults: a conceptual review. Psychiatry Res 2020 284 112732. (https://doi.org/10.1016/j.psychres.2019.112732)
- 10↑
Liu K, Chen W & Lin H. XG-PseU: an eXtreme Gradient Boosting based method for identifying pseudouridine sites. Mol Genet Genomics 2020 295 13–21. (https://doi.org/10.1007/s00438-019-01600-9)
- 11↑
Lundberg SM & Lee S-I. A unified approach to interpreting model predictions. In Proceedings of the 31st International Conference on Neural Information Processing Systems, pp 4768–4777. Eds SM Lundberg & S-I Lee. Red Hook, NY, USA: Curran Associates Inc, 2017. (https://dl.acm.org/doi/10.5555/3295222.3295230)
- 12↑
Kang D, Kim H, Cho J, et al. Prediction model for postoperative quality of life among breast cancer survivors along the survivorship trajectory from pretreatment to 5 years: machine learning-based analysis. JMIR Public Health Surveill 2023 9 e45212. (https://doi.org/10.2196/45212)
- 13↑
Bartley GB & Gorman CA. Diagnostic criteria for Graves’ ophthalmopathy. Am J Ophthalmol 1995 119 792–795. (https://doi.org/10.1016/s0002-9394(14)72787-4)
- 14↑
Bartalena L, Kahaly GJ, Baldeschi L, et al. The 2021 European Group on Graves’ orbitopathy (EUGOGO) clinical practice guidelines for the medical management of Graves’ orbitopathy. Eur J Endocrinol 2021 185 G43–G67. (https://doi.org/10.1530/eje-21-0479)
- 15↑
Lam AKC, Lam C, Leung W, et al. Intra-observer and inter-observer variation of Hertel exophthalmometry. Ophthalmic Physiol Opt 2009 29 472–476. (https://doi.org/10.1111/j.1475-1313.2008.00617.x)
- 16↑
Freedman K, Ray C & Kirk D. Reevaluation of current prism standards with recommendations to increase accuracy in the measurement of strabismus. Am J Ophthalmol 2019 198 130–135. (https://doi.org/10.1016/j.ajo.2018.09.009)
- 17↑
Wang Q, Qiao W, Zhang H, et al. Nomogram established on account of Lasso-Cox regression for predicting recurrence in patients with early-stage hepatocellular carcinoma. Front Immunol 2022 13 1019638. (https://doi.org/10.3389/fimmu.2022.1019638)
- 18↑
Lee YW, Choi JW & Shin E-H. Machine learning model for predicting malaria using clinical information. Comput Biol Med 2021 129 104151. (https://doi.org/10.1016/j.compbiomed.2020.104151)
- 19↑
Zhang Z, Zhao Y, Canes A, et al. Predictive analytics with gradient boosting in clinical medicine. Ann Transl Med 2019 7 152. (https://doi.org/10.21037/atm.2019.03.29)
- 20↑
Luo X, Wen X, Zhou M, et al. Decision-tree-initialized dendritic neuron model for fast and accurate data classification. IEEE Trans Neural Netw Learn Syst 2022 33 4173–4183. (https://doi.org/10.1109/tnnls.2021.3055991)
- 21↑
Albaradei S, Thafar M, Alsaedi A, et al. Machine learning and deep learning methods that use omics data for metastasis prediction. Comput Struct Biotechnol J 2021 19 5008–5018. (https://doi.org/10.1016/j.csbj.2021.09.001)
- 22↑
Al-Barakati HJ, Saigo H, Newman RH, et al. RF-GlutarySite: a random forest based predictor for glutarylation sites. Mol Omics 2019 15 189–204. (https://doi.org/10.1039/c9mo00028c)
- 23↑
Bradley EA, Sloan JA, Novotny PJ, et al. Evaluation of the national eye institute visual function questionnaire in Graves’ ophthalmopathy. Ophthalmology 2006 113 1450–1454. (https://doi.org/10.1016/j.ophtha.2006.02.060)
- 24↑
Jensen AL & Harder I. The impact of bodily change on social behaviour in patients with thyroid-associated ophthalmopathy. Scand J Caring Sci 2011 25 341–349. (https://doi.org/10.1111/j.1471-6712.2010.00832.x)
- 25↑
Zeng P, Fan S-X, Li Z-J, et al. Evaluation of the Graves’ orbitopathy-specific quality of life questionnaire in the mainland Chinese population. J Ophthalmol 2019 2019 1–5. (https://doi.org/10.1155/2019/7602419)
- 26↑
Hoppe E, Lee ACH, Hoppe D, et al. Predictive factors for changes in quality of life after steroid treatment for active moderate-to-severe Graves’ orbitopathy: a prospective trial. Eur Thyroid J 2020 9 313–320. (https://doi.org/10.1159/000508071)
- 27↑
Perros P, Crombie AL, Matthews JN, et al. Age and gender influence the severity of thyroid-associated ophthalmopathy: a study of 101 patients attending a combined thyroid-eye clinic. Clin Endocrinol 1993 38 367–372. (https://doi.org/10.1111/j.1365-2265.1993.tb00516.x)
- 28↑
Smith TJ, Cockerham K, Lelli G, et al. Utility assessment of moderate to severe thyroid eye disease health states. JAMA Ophthalmol 2023 141 159. (https://doi.org/10.1001/jamaophthalmol.2022.3225)
- 29↑
Ponto KA, Pitz S, Pfeiffer N, et al. Quality of life and occupational disability in endocrine orbitopathy. Dtsch Arztebl Int 2009 106 283–289. (https://doi.org/10.3238/arztebl.2009.0283)
- 30↑
McBain HB, MacKenzie KA, Au C, et al. Factors associated with quality of life and mood in adults with strabismus. Br J Ophthalmol 2014 98 550–555. (https://doi.org/10.1136/bjophthalmol-2013-304220)
- 31↑
Du B, Wang Y, Yang M, et al. Clinical features and clinical course of thyroid-associated ophthalmopathy: a case series of 3620 Chinese cases. Eye 2021 35 2294–2301. (https://doi.org/10.1038/s41433-020-01246-7)
- 32↑
Wang Y, Sharma A, Padnick-Silver L, et al. Physician-perceived impact of thyroid eye disease on patient quality of life in the United States. Ophthalmol Ther 2021 10 75–87. (https://doi.org/10.1007/s40123-020-00318-x)
- 33↑
Choi YJ, Lim HT, Lee SJ, et al. Assessing Graves’ ophthalmopathy-specific quality of life in Korean patients. Eye 2012 26 544–551. (https://doi.org/10.1038/eye.2011.359)
- 34↑
Lin I-C, Lee C-C & Liao S-L. Assessing quality of life in Taiwanese patients with Graves’ ophthalmopathy. J Formos Med Assoc 2015 114 1047–1054. (https://doi.org/10.1016/j.jfma.2013.12.002)
- 35↑
Stiebel-Kalish H & Robenshtok E. Thyroid eye disease: therapy in the active phase: a comment. J Neuroophthalmol 2014 34 425–426. (https://doi.org/10.1097/wno.0000000000000187)
