This is a multicentre and pilot investigation. The study included all patients from the Catalan Institute for Fostering and Adoption (ICAA) database, who agreed to participate. The total study cohort comprised 231 patients, which includes 73 control patients and 158 patients diagnosed with FASD. The study, registered at clinicaltrials.gov (NCT02558933), integrated cohorts from previous investigations (PI13/01135; OG085818; PI16/00566; PI19/01853) that included participants enrolled between March 2017 and November 2023. The study was conducted at the Hospital del Mar Medical Research Institute of Barcelona and Hospital Clinic of Barcelona, and all procedures adhered to ethical standards outlined in the Declaration of Helsinki and Spanish data privacy regulations. Consent was obtained from the caregiver or legal representative of patients due to their incapacity to provide informed consent, as approved by the Comité Ético de Investigación Clínica Parc de Salut MAR (No. HCB/2021/0459).

The minimum sample size calculation was conducted using G^*Power software (Faul et al., 2007) with the following parameters: bilateral contrast, alpha 0.1, beta cut-off of 0.2, corresponding to a power of 0.8 (Gupta et al., 2016), estimated proportion of replacements required (20%), precision 0.1 (90%) and estimate 50% of the population affected. A minimum of 70 samples in two independent groups (non-FASD and FASD groups or Non-FASD vs. each FASD subtype) were required to compute the necessary sample size.

To diagnose FASD, all adopted children in the EEC who were included in this study, including those with verified prenatal alcohol exposure (PAE), underwent independent examination using standardized dysmorphology exams (Hoyme et al., 2016). The diagnostic category for each child was identified based on the 1996 IOM standards (reviewed in 2016) (Hoyme et al., 2005, 2016), which consist of five diagnostic characteristics. (1) confirmed prenatal alcohol exposure; (2) evidence of a characteristic minor facial abnormalities pattern, typified by having a thin upper lip, smooth philtrum and short palpebral fissures; (3) growth retardation, defined as height or weight ≤ 10th percentile; (4) evidence of deficient brain growth or subrogated data; and (5) behavioral or cognitive affected domains (1 or 2) related to prenatal alcohol exposure. For a diagnosis of complete FAS, criteria 2, 3, 4, 5 (confirmed or not confirmed prenatal alcohol exposure) were required. For partial FAS, criteria 1, 2, and at least one of criteria 5 (confirmed prenatal alcohol exposure) or 2, 5 and 3 or 4 (no confirmed prenatal alcohol exposure) were required. The diagnosis of alcohol-related birth defects (ARBD) required the finding of one criterion plus a minimum of one structural defect involving heart, skeleton, kidney, eye, ear or minor abnormalities like railway ears, midface hypoplasia or stick hockey hands. The diagnosis of alcohol-related neurodevelopmental disorders (ARND) required the finding of 1 and 5 criteria.

The variables selected for our study were based on IOM classification from 2016 (Hoyme et al., 2016) to FASD diagnosis. Selected sociodemographic variables are related to maternal alcohol consumption during pregnancy (Hoyme et al., 2016), origin and ethnicity (Oh et al., 2023a). Parent feeling variables were included in our selection as they may be relevant to determine if parental perceptions play a role in the diagnosis of FASD predicted by ML algorithms. For clinical variables were selected growth deficits (Astley et al., 2016; Hoyme et al., 2016; Treit et al., 2016), craniofacial dysmorphology (Smith et al., 2014; Hoyme et al., 2016), birth malformations (Dylag et al., 2023), neurodevelopmental disorders (Geier and Geier, 2022) and other physical features and medical history (Brennan and Giles, 2014; del Campo and Jones, 2017; Ninh et al., 2019). Lastly, related to neuropsychological domains, we selected variables significant for FASD diagnosis, including motor cognition (Bakoyiannis et al., 2014), language (Hendricks et al., 2019), academic achievement (Glass et al., 2017), memory (Rasmussen, 2005), attention (Young et al., 2016), executive functioning including impulse control and hyperactivity (Peadon and Elliott, 2010), affect regulation (Temple et al., 2019) and adaptive behavior, social skills, or social communication (Temple et al., 2019; Hammond et al., 2022).

Cognitive assessment utilized the Wechsler Intelligence Scale for Children (WISC) series, using the Fifth Edition (WISC-V) (Weiss et al., 2019). Evaluating full-scale Intelligence Quotient (IQ), Verbal Comprehension (VCI), Visuospatial Index (VSI), Perceptual Reasoning (FRI), Working Memory (WMI), and Processing Speed (PSI).

Adults' cognitive functioning was evaluated using the Wechsler Adult Intelligence Scale (WAIS), the Fourth Edition (WAIS-IV) (Wechsler, 2008). Preschoolers' cognitive abilities were assessed using the Wechsler Preschool and Primary Scale of Intelligence (WPPSI-IV) (Raiford and Coalson, 2014). Additionally, the Adult Self-Report (ASR/18-59) (Achenbach and Rescorla, 2003) collected self-reported data on behavioral concerns in adults, while the Child Behavior Checklist (CBCL) (Achenbach, 2004) gathered parental reports on children aged 6–18. All assessments adhered to unified criteria and professionals received standardized training, ensuring consistency and reliability. Data were recorded in a confidential database, maintaining accuracy and confidentiality throughout the research process.

Statistical analysis was performed using SPSSv22 and R. Graphs were performed using Graphpad Prism 8.0 software. Descriptive analysis was used to characterize the samples. Categorical variables were presented as counts and percentages, while continuous variables were presented as means and standard deviations. Relationships between sociodemographic, clinical and neuropsychological features were examined for quantitative factors using Kruskall–Wallis test with Dunn's correction for multiple comparisons and for qualitative variables chi-square test. A significance level of p < 0.05 was applied to all analyses.

In addition to the aforementioned statistical tests, machine learning (ML) algorithms were also employed to create a predictive model, using the statistical software R (3.3.0+ version).

This study employed several ML algorithms to predict FASD and its subtypes (FAS, pFAS and ARND), such as LR, LDA, linear SVM, polynomial SVM, KNN, RF and XGB (Zhang et al., 2019). The data underwent a preparation phase, where “mice” function from VIM package was used for missing values imputation, employing the predictive mean matching method (pmm) (Kowarik and Templ, 2016). This process was repeated 5 times, as per default setting. Just 1% of the data were missing and single imputation is considered appropriate when <5% of the data are missing (Graham, 2009). The dataset was subsequently scaled using “scale” function in base R.

LR is a binary classification algorithm, which uses a logistic function to predict class probability. Coefficients are obtained using maximum likelihood estimation (Hosmer et al., 2013). LR is easy to implement and performs well on linearly separable classes. However, it may overfit with many features and struggles with complex relationships. LDA projects data into a lower-dimensional space, maximizing class separability and minimizing variance within a class, finding a linear combination of features that characterizes a group (Gardner-Lubbe, 2021). SVM maximizes the distance between the separating hyperplane of the variables to classify (Huang et al., 2018). In our study, linear SVM and polynomial SVM are differentiated. Linear SVM classifies linearly separable data and is more computationally efficient. Polynomial SVM classifies non-linearly separable data, transforms input space into a higher-dimensional space, finds more complex relationships, and is computationally intense (López et al., 2022). KNN predicts class by calculating the Euclidean distance to all training points and selecting K most similar instances (the neighbors). It handles multiclass classification and learns complex decision boundaries (Zhang, 2016). However, it performs poorly on high-dimensional datasets because the distance to all neighbors must be recalculated. Ensemble methods like RF and XGB are decision tree-based algorithms. RF combines multiple independently trained decision trees, uses bagging to create subsets of the original dataset, and then aggregates the results (Denisko and Hoffman, 2018). On the other hand, XGB trains decision trees sequentially, with each new tree correcting errors made by the previous one (Li et al., 2022). The ML algorithms used in this study has its own strengths and weaknesses, leading to varied results. The range of ML algorithms compared spans from traditional predictive models like LR to more complex ensemble methods like RF and XGB, which are capable of handling high-dimensional data. By comparing different models, our study aimed to find the most effective model for predicting FASD and its subtypes. This diversity in approaches enhances the robustness and comprehensiveness of the study.

For the analysis, a total of 66 variables were selected, encompassing five sociodemographic parameters, 35 clinical features, six intelligence scores and 20 behavioral domains (Tables 1–4). Prior to model construction, a hold-out method was applied to split the data into training and test sets using “createDataPartition” function from caret package in R (Kuhn, 2008). Sixty-seven percent of the data was allocated to training set and the remaining 33% to test set. This function employs a stratified random sampling method, which minimizes the bias of the data distribution and creates balanced data.

In addition to the hold-out method, a resampling method involving four-fold cross-validation and three repeats was adopted. This was implemented using “trainControl” function from the caret package (Kuhn, 2008). The models were trained using “train” function with hyperparameters set to default, which gathers and simplifies numerous R algorithms for the development of predictive models (Kuhn, 2008). The models employed included LR, using “glm” method and binomial family, and LDA, implemented with “lda” method, which has “moment” as the default mean and variance estimator. Linear SVM and Polynomial SVM were performed using “svmLinear” and “svmPoly” methods, respectively. They have C tuning parameter, which determines the margin classification, equal to 1 as default settings. KNN was employed by “knn” method also from caret package, performing automatic hyperparameter tuning for k depending on instance-based learning. In addition, RF was employed using “rf” method, with 500 trees as default. XGB model used “xgbTree” method, having 100 maximum iterations by default.

The “predict” function from stats package was used to predict classes with the test group. In order to make comparisons, the “confusionMatrix” function from caret package was used to calculate true positive, true negative, false positive and false negative. These calculations provided measures including accuracy, precision, sensitivity, F1 score and specificity. ROC-AUC was obtained using “roc” function from pROC package (Robin et al., 2011). Training and test datasets were consistent across FASD and its subgroups, ensuring a fair and valid comparison.

Feature importance prediction of the models was determined by calculating the Root Mean Square Error (RMSE) loss after permutation. It was obtained with “explain” function from DALEX package, with “classification” type model in arguments (Law Biecek, 2018). Plots were generated from the object class formed by “variable_importance” function from caret package (Kuhn, 2008).

The study initially included 273 patients. However, 42 were excluded: 28 lacked psychological evaluations and 14 refused participation. Of the remaining 231 subjects, 73 were diagnosed as non-FASD (controls), and 158 were diagnosed with FASD, comprising 33 with FAS, 81 with pFAS, and 44 with ARND. A database was compiled with sociodemographic and psychological characteristics of both FASD patients (and their subtypes) and non-FASD participants (Figure 1).

Sociodemographic characteristics of the population were collected from FASD and non-FASD patients (Table 1). The chi-square test revealed no significant differences between groups. Significant differences were found in physical characteristics between children diagnosed with FASD and non-FASD (Table 2, Supplementary Table 1). Prematurity (p-value = 0.011) was higher in FAS children compared to non-FASD children (p-value = 0.018; Supplementary Table 1). Growth retardation (p < 0.001) was also higher in FAS and ARND children compared to non-FASD children (p-value <0.001; p-value = 0.040). Birth complications (p-value = 0.047), as perinatal asphyxia or abnormal heart rate, were more prevalent in ARND children compared to non-FASD children (p-value = 0.052). As expected, maternal alcohol consumption confirmation also showed significant differences (p-value <0.001) in all groups compared to non-FASD patients.

FASD patients showed lower height (p < 0.0001) by 81%, 16%, and 25% for FAS, pFAS, and ARND respectively, compared to non-FASD (Table 2, Supplementary Table 1). Weight alterations (p < 0.0001) increased significantly in FAS and pFAS. Microcephaly (p-value <0.0001), shorter palpebral fissures (p-value = 0.01) and lip-philtrum affectation (p-value <0.001) were more prevalent in FAS and pFAS. Facial anomalies (p-value = 0.025) were significantly higher in pFAS (p-value = 0.009) compared to non-FASD group. In particular, children with affected eyes (p-value = 0.013) and upper limbs (p = 0.001) were predominantly from FAS groups compared to non-FASD (p-value = 0.037 and p-value = 0.001, respectively). Moreover, significant disparities in lower limbs (p-value = 0.034) were observed, primarily in ARND compared to non-FASD group.

Among FASD groups (Table 2, Supplementary Table 1), significant differences included prematurity (p-value = 0.009), eyes (p-value = 0.011) and upper limbs affectation (p-value = 0.011) between FAS and ARND. FAS exhibited higher growth retardation levels compared to pFAS (p-value = 0.014). Maternal alcohol consumption, short palpebral fissures and lip-philtrum affectation showed increased levels in FAS (p-value = 0.042, p-value = 0.003 and p-value <0.001, respectively) and pFAS patients (p-value = 0.023, p-value = 0.018 and p-value <0.001, respectively) compared to ARND, confirming that this group does not exhibit physical characteristics. Microcephaly varied among all FASD groups, showing 87% of the cases in FAS, 46% in pFAS and 18% in ARND. No significant differences were noted in other physical or clinical characteristics, except for a trend toward greater cardiac damage in FASD (p-value = 0.06).

Significant differences were observed in psychological intelligence parameters between FASD and non-FASD patients. Evaluating cognitive performance with WISC V, children diagnosed with FASD, specifically FAS and pFAS groups (Table 3, Supplementary Table 2), exhibited lower scores on VCI (p-value = 0.001), VSI (p-value = 0.005), FRI (p-value = 0.001), WMI (p-value <0.001), PSI (p-value = 0.003) and IQ (p-value <0.001). No significant differences were found for the WAIS-IV and WPPSI-IV tests.

Related to behavioral parameters, CBCL 6–18 test showed impairments in several cognitive domains in FASD patients compared to their non-FASD counterparts (Table 4, Supplementary Table 3). Increased levels of thought problems (p-value = 0.035), rule breaking behavior (p-value = 0.002), externalizing problems (p-value = 0.045), total problems (p-value = 0.008), anxiety problems (p-value = 0.009), obsessive compulsive problems (OCP; p-value = 0.049) and stress problems (p-value = 0.001) domains were observed in ARND compared to non-FASD. Significantly increased levels of attention problems were observed in FAS (p-value = 0.021) and ARND (p-value = 0.020) compared to non-FASD. Within the FASD group found significant differences in thought problems (p-value = 0.007), anxiety problems (p-value = 0.006) and OCP (p-value = 0.003), with significantly increased levels in the ARND group compared to pFAS. Furthermore, ARND showed high levels of rule-breaking behavior compared to FAS and pFAS subgroups (p-value = 0.002 and p-value = 0.002), externalizing problems (p-value = 0.003 and p-value = 0.002), total problems (p-value = 0.006 and p-value = 0.021), oppositional defiant problems (p-value = 0.002 and p-value = 0.018), conduct problems (p-value = 0.016 and p-value = 0.009) and stress problems (p-value <0.001 and p-value = 0.015), respectively. Moreover, the results display significantly increased levels of aggressive behavior (p-value = 0.01) in the ARND group compared to FAS.

Finally, in the adult behavioral test ASR 18–59 (Table 4, Supplementary Table 3), significant differences related to attention problems (p-value = 0.032) were observed, showing ARND higher levels compared to the pFAS subgroup (p-value = 0.003).

Predictive models for FASD diagnosis were developed using ML, considering the sociodemographic, clinical, and psychological variables previously discussed. The dataset consisted of 231 samples, with 155 samples used for model training and the remaining 76 samples saved for testing and final model evaluation. A variety of ML algorithms were employed, including XGB, LR, LSVML, LDA, SVMP, kNN, RF and XGB. These models were trained using four-fold cross-validation on the training dataset.

Figure 2 shows the key performance metrics associated with the predictive power of each model. Among the models, the ensemble algorithms (RF and XGB) outperformed the others. Notably, the RF model achieved the highest accuracy (0.92), precision (0.96), sensitivity (0.92), F1 score (0.94), specificity (0.92), and AUC (0.92), establishing it as the most effective model for predicting FASD diagnosis. Other models such as LR, LDA, SVMP, and kNN showed lower performance on these metrics (Figure 2). Consequently, we selected the RF model for our prediction tasks due to its superior discriminative ability.

To understand the decision-making mechanism of the RF model, we examined the significance of the variables within this algorithm. The features were ranked according to their importance, with maternal alcohol consumption being the most significant (0.48), followed by lip-philtrum (0.27), microcephaly (0.19), height affectation (0.17), Working Memory Index (0.16), aggressive behavior (0.16), Intelligence Quotient (0.15), somatic complaints (0.15), weight affectation (0.15), and depressive problems (0.15; Figure 3). These findings offer crucial insights into the primary attributes associated with FASD conditions and their respective significance in the predictive model.

Another aim of our study is to construct individualized models for each category of FASD. This methodology will allow us to uncover distinct attributes and trends that might remain concealed when all FASD types are examined collectively.

Focusing our analysis on FAS prediction in comparison to non-FASD, we employed the previous ML algorithms. The XGB model outperformed the others (Figure 4), achieving the highest accuracy (0.94), precision (0.91), sensitivity (0.91), F1 Score (0.91), specificity (0.96), and AUC (0.93), thereby proving to be the most effective model for FAS diagnosis prediction.

In Figure 5 the features were ranked based on their importance, with Height (0.32) and Weight (0.28) being the most influential, followed by Fluid Reasoning Index (0.11), Internalizing Problems (0.08), Total Problems (0.1), and Processing Speed Index (0.1) (Figure 5). This highlights FAS prediction is mainly determined by failure to thrive.

In the subsequent stage of the research, the focus shifted to the prediction of pFAS compared to non-FASD. Upon evaluating all ML models (Figure 6), the RF model emerged as the most proficient, achieving the highest metrics in accuracy (0.90), precision (0.86), sensitivity (0.96), F1 Score (0.91), specificity (0.83), and AUC (0.90). This underscores its effectiveness in predicting pFAS diagnosis.

Lip-philtrum (0.36) and Maternal Alcohol Consumption (0.27) were the most impactful features for pFAS prediction, followed by Intelligence Quotient (0.21), Microcephaly (0.18), and Processing Speed Index (0.16), Verbal comprehension index (0.15), attention problems (0.15) and thought problems (0.15; Figure 7).

The final phase of the study involved the analysis of ARND prediction. As previously observed, the RF model demonstrated superior performance for ARND prediction (Figure 8), obtaining the best levels of accuracy (0.87), precision (0.76), sensitivity (0.93), F1 Score (0.84), specificity (0.83), and AUC (0.88).

Figure 8 displays Maternal Alcohol Consumption (0.64) as the most influential feature for ARND prediction, followed by Total Problems (0.11) and Attention Problems (0.10) (Figure 9). These results confirm the importance of PAE confirmation for diagnosis.