Construction and implications of structural equation modeling network for pediatric cataract: a data mining research of rare diseases

A total of 237 patients registered with the CCPMOH [11] were recruited from Zhongshan Ophthalmic Center [12], one of China’s largest eye hospitals, located in Guangzhou city, south China. Seventy-seven patients with wide range of incomplete or missing clinical record were excluded. All the included patients are diagnosed with pediatric cataract according to the International Statistical Classification of Diseases and Related Health Problems 10th Revision (ICD-10) Version (Disease numbers: H26.0 and Q12.0) [12]. Finally, 160 patients’ data records were included into the analysis and SEM construction. The dataset was anonymized throughout the research.

The research protocol was approved by the Institutional Review Board/Ethics Committee of Sun Yat-sen University (Guangzhou, China). Informed written consent was obtained from at least one family member of each participating patient, and the tenets of the Declaration of Helsinki were followed throughout this study. To allow confidential evaluation of the use of a slit-lamp, a Tono-Pen, a Pentacam imaging system and the Teller visual acuity (VA) cards in this study, this trial was registered with the Clinical Research Internal Management System of CCPMOH. The authors affirm that all ongoing trials and trials related to this study are registered.

All original variables and their SEM constructions are displayed in Table 1.

The age at diagnosis, height, weight, family hereditary history, abnormal parturition history, and abnormal pregnancy history were collected in regular clinical practice. The diagnosis of pediatric cataract was made by experienced ophthalmologists according to the ICD-10 Version [13]. The family hereditary history was defined as any similar disease history of immediate family members. The abnormal parturition history included but was not limited to an abnormal fetal position, abnormal placenta and amniotic fluid, fetal distress, breech birth, fetal macrosomia, and hypamnion. The abnormal pregnancy history included premature delivery, post-term pregnancy, pregnancy complicated with infection, gestational hypertension and gestational diabetes mellitus.

Area, density, and location are three critical lesion indices that are defined for the comprehensive evaluation and treatment decisions of pediatric cataracts (details definition in Additional file 1) [14].

Laterality was classified as bilateral cataracts or a unilateral cataract. For unilateral cataracts, data from the affected eye were included in the analysis and modeling. For bilateral cataracts, if the interocular lesion (area, density, and location) presented no differences, only the variables of the right eye were chosen for the modeling and analysis; otherwise, data from this patient were excluded to avoid bias caused by reduplicative datasets.

The ocular complications consisted of microphthalmia, micro- or megalocornea, keratoconus, glaucoma, traumatic or complicated cataracts, or vitreous and retinal diseases. Otherwise, the data were regarded as “no ocular complications”.

The intraocular pressure (IOP) measurement was conducted using a Tono-pen tonometer (Reichert Inc., Seefeld, Germany) (details in Additional file 1) [15, 16].

For the pre-modeling, the age at diagnosis, height, weight, family hereditary history, abnormal parturition history, and abnormal pregnancy history were considered the concomitant variables. The AL, area, density, and location were loaded into the one-layer latent variance and termed the structural index (STR). The ocular complications, IOP, UCVA and BCVA were loaded into another one-layer latent variance and termed the functional index (FUN). The overall index (OVE) was settled as the two-layer latent variance. The aforementioned structure for pre-modeling is presented in Table 1.

In the pre-modeling process, primary significance analysis and iteration testing were applied for variable fitting. The combination of the variables for the best SEM fitting will be accept and proceeded into further analysis. After pre-modeling, a total of 8 filtered variables were selected and included in the final SEM (Table 2).

SEM was conducted to statistically test the interrelationships of the constructs and their relationships with the structural, functional, and overall components in our study population. Prior to modeling the relationship between latent variables, a measurement model was evaluated for each dietary behavior component. This step involves a confirmatory factor analysis to reveal the relationship between the latent variables and their indicator variables. The following step (i.e., testing of the structural model) estimates the strength of the relationships between these latent variables. It also allows for an examination of the direct and indirect effects among the constructs in the model. The data were examined prior to modeling to ensure that they met the assumptions of the proposed SEM and were analyzed using the robust weighted least squares procedure in Mplus 7.0 (Mplus, Los Angeles, Calif) [21].

To evaluate the goodness of fit of the model, the χ²-value was calculated together with the degrees of freedom and four other indices as follows: the root mean square error of approximation (RMSEA), the Tucker-Lewis index (TLI), the comparative fit index (CFI), and the weighted residual root mean square residual (WRMR). Values below 0.08 for RMSEA, below 1.0 for WRMR and above 0.90 for TFI and CFI indicate an acceptable fit of the data to the hypothesized model (details in Additional file 1).

A total of 160 participants from the CCPMOH database with 8 filtered variables were included in the final SEM analysis (details in Table 2). The mean age of the included participants was 50.99 months ± 36.38 months and 23.13% (n = 37) of our patients had an abnormal pregnancy history. For the STR network, the mean AL value was 21.75 mm ± 2.06 mm, 55% (n = 88) of our patients had an extensive area, 37.5% (n = 60) of our patients had dense opacity, and 60.63% (n = 97) of our patients’ opacities had central locations. For the FUN network, the mean IOP value was 15.12 ± 6.88 mmHg, and the mean UCVA value (logMAR) was 0.77 ± 0.44.

As shown in Table 3, the overall fit of the measurement and the structural models was satisfactory (χ² = 26.093, df = 16, P = 0.1113; RMSEA = 0.053, P = 0.422; CFI = 0.995; TFI = 0.992; WRMR = 0.781). Finally, the estimates of the SEM constructs are presented in Table 4.

The standardized SEM architecture and interrelationships of the constructs are displayed in Fig. 2.

The results from SEM network quantifiably illuminate the effect of eight potential factors (four risk factors: area, density, location, and abnormal pregnancy experience; four beneficial factors: AL, UCVA, IOP value, and age at diagnosis). For structural effect, extensive lesion area, dense opacity and central location will present increasing severity of cataract but AL growth will be a positive indicator for the better structural maturity. Furthermore, both UCVA and IOP value in normal range presented beneficial functional effect and could be the positive evaluation indicators for patients. Additionally, patients with earlier diagnosis age and abnormal pregnancy experience will be the indicators for a more severe status. All these identified factors are also valuable references and indicators for patients’ prognoses.

Studying rare and complex disease is a great challenge for researchers and medical practitioners [22]. Rare and complex diseases tend to be caused by a system, with interconnected entities being modeled as nodes and their connections as edges to comprise an intricate pathogenic network [23]. Therefore, it is always difficult for researchers to explicitly elucidate the principle nosogenesis, individual predisposition and related risk factors.

With the advance and popularization of medical detecting and monitoring equipment, medical practitioners should address an increasing number of clinical indices. However, quite a few examinations or clinical indices seem to be unnecessary or to provide ambiguous information for the diagnosis and the treatment decision, which is contrary to the goals of precision medicine [24, 25]. Therefore, it is imperative to uncover the interrelationship and the effectiveness of these potential clinical indices and risk factors and to integrate these elements to construct a holistic network capable of providing insights into rare disease evidence-based prevention and treatment.

To obtain sufficient evidence and data for rare diseases, we established the CCPMOH program and built a rare disease data integration platform [11]. Medical history, multiple structural indices, and comprehensive functional metrics were collected for the primary investigation. After pre-modeling, a total of 8 filtered variables were selected and included in the final SEM modeling.

Rare diseases, including pediatric cataracts, are always associated with birth defects [26]. However, pediatric cataracts present great genetic heterogeneity, and a considerable proportion of pediatric cataracts should be considered in the context of interactions among the environment and heredity [27]. Among the 8 filtered variables, abnormal pregnancy history was included but family hereditary history and abnormal parturition history were excluded from the final network. In addition to the final network contribution, these results indicate that abnormal pregnancy history may be the principle risk factor for pediatric cataracts. Previous studies reported that pediatric cataracts had an association with rubella infection [28, 29], which was in agreement with our findings on the abnormal pregnancy history. Therefore, more attentions should be focused on patients with abnormal pregnancy history in screening and early diagnostics of pediatric cataract.

To systematically model our clinical indices, we classified our records into two separated latent variances (structural indices or function indices) for two-layer SEM construction. Two essential steps are needed for the SEM clinical interpretation. First, we should holistically evaluate the SEM network using existing evidence and the clinical consensus. Our results indicate that the functional and structural networks have positive support from the overall network and that the structural network has a positive supporting effect on the functional network. Additionally, the lesion area, density and location have a negative effect on the structural network, and UCVA is a positive indicator for the functional network. These results are concordant with conventional medical knowledge and validate the reliability of the SEM network.

After verifying the reliability of the model in the preliminary test, the second step is to explain the model and provide novel and valuable clinical information. For the structural network, AL presents a positive effect in our study population. The main study population was younger than 2 years of age; thus, in the early stage, increasing AL is a beneficial indictor for eyeball development instead of myopia progression. For the functional network, the IOP plays as a supporting role, which reveals that an IOP in the normal range (the main range of our IOP is below 21 mmHg) is not a risk factor for glaucoma but instead is a supporting factor for visual function and a positive nutrient media for the functional component. For the overall network, age at diagnosis is also a positive indicator, which demonstrates that the severe pediatric cataract is more likely to be diagnosed in the early stage. These implications could be an important supplement to the mainstream theory of pediatric cataract.

The weight of each index is also valuable information for the clinical interpretation. For the structural network, AL and density are the dominant factors and can be considered key indicators for severity and prognosis. For the functional network, UCVA plays the main role along with the subordinate beneficial effect from the IOP. The functional and structural networks almost equally provide a contribution to the overall network. Furthermore, the age at diagnosis is more crucial than an abnormal pregnancy experience in the overall network. Therefore, AL, density, UCVA and age at diagnosis might be emphasized in our regular clinical practice.

We performed the goodness-of-fit validation to ensure the meaning and reliability of the SEM network. The key indicators, χ²-values and RMSEA indicated an acceptable fit of the data to our model along with the TFI, CFI, and WRMR, which supported the same conclusion. Although the modeling reliability of our variables was not sufficiently ideal to show great explanation power, the overall fitting of the SEM network is acceptable.

The main limitation of our study is the data quality. Two main effects (missing data and inevitable errors) caused by data quality are introduced as follows. Missing data will directly influence the model fitting; thus, the reason for the exclusion of the primary indices should not be merely interpreted by their limited or subordinate effect on disease. Notably, some valuable indices might be excluded by the effect of the missing data. Moreover, the missing data could be a possible explanation for the unsatisfactory modeling reliability. Additionally, uncooperative measurement is common for our regular examination procedure because the majority of our patients are under 2 years of age, which causes fairly inevitable measuring errors. Furthermore, the VA evaluation is a somewhat subjective measurement and the medical history is a retrospective inquiry, which might present less accuracy. Our future study aims are to include a more dimensional and more complete rare disease dataset with a larger sample size for modeling and data mining.