Large language models for generating longitudinal synthetic data in low-risk pregnancy care Authors: Gonçalo Pinto, Vítor Crista, Regina Silva, Bruno Costa, Filipe Andrade Bernardi, Rute Almeida, Julio Souza, Alberto Freitas, Goreti Marreiros, Luís Conceição Published: 2026 ABSTRACT AbstractThis article describes an obstetric dataset covering the full continuum of care of 5000 synthetic low-risk pregnant women, from preconception to post-birth follow-up, generated using a large language model with zero-shot prompting. It includes trimester-specific clinical measurements, such as gestational weight, hemoglobin levels, glucose testing, anemia markers, and body mass index, as well as variables intrinsic to perinatal care, including preconception characteristics, antenatal consultations, and delivery outcomes.The dataset was generated through iterative prompt engineering. An initial prompt defined the longitudinal schema, essential variables, and domain-specific generation rules. The output was assessed by an expert panel in the field of General and Family Medicine and Data Science. Successive refinements incorporated expert feedback to improve plausibility, auditability, and overall consistency. After generation, the dataset underwent rule-based and plausibility-oriented validation procedures focused on internal statistical coherence and clinical assumptions’ assurance.In addition to mitigating data access constraints and privacy issues, the dataset enables the reconstruction of longitudinal obstetric care pathways for simulation studies. It can also be reused for methodological development, benchmarking, and educational machine learning experiments. However, it was not designed to directly support clinical decision-making and clinical model deployment.Keywords: Obstetrics, Synthetic patients, Artificial intelligence, Tabular data, Healthcare Specifications Table SubjectHealth Sciences, Medical Sciences & PharmacologySpecific subject areaSynthetic healthcare data generation for longitudinal low-risk pregnancy care using large language models.Type of dataTableData collectionSynthetic patient-level pregnancy trajectories were generated using a large language model (ChatGPT web interface) in a zero-shot setting. Prompts defined essential variables based on the Portuguese National Low-Risk Pregnancy Surveillance Program and encoded domain rules including visit patterns, clinical and outcome ranges, and mandatory causal relationships. No real patient data was used. Data generation was restricted to low-risk profiles. ChatGPT output included the synthetic dataset in tabular format (CSV/XLSX), a data dictionary, clinical rules and assumptions, and limitations. Post-processing included validation procedures to assess data quality, statistical coherence, and clinical plausibility.Data source locationInstitution: School of Engineering – Polytechnic Institute of Porto (ISEP)City/Town/Region: Porto, Porto districtCountry: PortugalData accessibilityRepository name: ZenodoData identification number: 10.5281/zenodo.20204557Direct URL to data: https://zenodo.org/records/20204557Related research articleNone. Value of The Data •The dataset provides a structured synthetic longitudinal resource for low-risk pregnancy care in settings where regulatory and ethical constraints limit access to real-world data. The dataset provides a structured synthetic longitudinal resource for low-risk pregnancy care in settings where regulatory and ethical constraints limit access to real-world data. •The dataset’s longitudinal structure and clinically constrained trajectories enable simulation studies and the construction of machine learning models for methodological development, benchmarking and teaching purposes. It also supports predictive tasks related to pregnancy outcomes, in particular cesarean delivery, gestational diabetes and hypertension, preeclampsia, preterm birth, and macrosomia. The dataset’s longitudinal structure and clinically constrained trajectories enable simulation studies and the construction of machine learning models for methodological development, benchmarking and teaching purposes. It also supports predictive tasks related to pregnancy outcomes, in particular cesarean delivery, gestational diabetes and hypertension, preeclampsia, preterm birth, and macrosomia. •Benchmarking outcome values can be evaluated using the dataset, namely: preterm birth rates ranging between 4-16%; recommended 10-15% benchmark cesarean delivery rates; share of pregnant women aged 35 or older at approximately 20%; prevalence rates of approximately 5% for gestational diabetes and between 5%-10% for gestational hypertension; preeclampsia reported in approximately 2-8% of pregnancies; and macrosomia prevalence estimated at approximately 12%, with ranges between 1.6% to 28% also clinically plausible. The dataset is accompanied by R scripts with examples of basic predictive modelling tasks. Benchmarking outcome values can be evaluated using the dataset, namely: preterm birth rates ranging between 4-16%; recommended 10-15% benchmark cesarean delivery rates; share of pregnant women aged 35 or older at approximately 20%; prevalence rates of approximately 5% for gestational diabetes and between 5%-10% for gestational hypertension; preeclampsia reported in approximately 2-8% of pregnancies; and macrosomia prevalence estimated at approximately 12%, with ranges between 1.6% to 28% also clinically plausible. The dataset is accompanied by R scripts with examples of basic predictive modelling tasks. •The synthetic dataset is accompanied by a structured, comprehensive data dictionary supporting data understandability and auditability, and a validation framework grounded on evidence-based rules and clinical assumptions constraints aligned with standard data quality dimensions, such as consistency and credibility. The synthetic dataset is accompanied by a structured, comprehensive data dictionary supporting data understandability and auditability, and a validation framework grounded on evidence-based rules and clinical assumptions constraints aligned with standard data quality dimensions, such as consistency and credibility. •It adds reproducible methodologies to support synthetic data generation research and validation strategies in the absence of ground truth or real-world comparator data. It adds reproducible methodologies to support synthetic data generation research and validation strategies in the absence of ground truth or real-world comparator data. Background Access to high-quality data is paramount for advancing research and developing Artificial Intelligence (AI) models in healthcare. However, real-world data are often difficult to obtain due to institutional barriers, privacy regulations, and ethical concerns [1,2]. Synthetic data generation has emerged as a privacy-preserving alternative by enabling artificial datasets that preserve key domain and statistical properties while mitigating ethical risks [3,4]. In healthcare, synthetic data supports simulation and prediction studies, hypothesis testing, methodologies and algorithm exploration, epidemiology and public health research, software development and testing, and education [5]. Recently, large language models (LLMs) have offered data generation capabilities in zero or few-shot settings by encoding domain-specific knowledge through prompting [6]. In fact, ChatGPT, a distinguished LLM, has demonstrated utility in synthetic healthcare data generation [7]. In this article, we present a dataset of longitudinal pregnancy care trajectories for 5,000 synthetic low-risk women generated using a large language model. Pregnancy care trajectories represent a suitable use case due to well-defined clinical pathways and established guidelines for low-risk populations [[8], [9], [10]]. Data Description This dataset is provided in tabular format and contains 5,000 synthetic patient records representing longitudinal low-risk pregnancy care trajectories, thereby not containing any real patient information. The complete dataset (Table 1) is publicly available as a single Excel file (XLSX extension) in both Portuguese and English, under the names “Complete Pregnant patients - Synthetic Dataset.xlsx” (English version) and “Complete Pregnant patients - Synthetic Dataset_portuguese.xlsx” (Portuguese version). Files can be freely downloaded from the Zenodo repository using the link provided in the specifications table. No request or authorization is required to access the data. Table 1Content of the public Excel file provided for use.Table 1 dummy alt textSheetContent (sheet name)Description1READMEOverview of dataset structure, sheet organization, and guidance for interpretation and reuse2Main dataset (dataset_en)Patient-level synthetic longitudinal pregnancy data of 5000 low-risk women. All variables are structured across specific clinical timepoints3Data dictionary (data_dictionary_en)Includes detailed information for each variable including name, definition, type, units/expected ranges accompanied by literature references, allowed values, timepoint of collection within the clinical pathway, and derivation logics specifying base variables and additional variables required for auditing.4Validation report (validation_summary)Summary of rule-based validation procedures and consistency checks applied, including the number of violations detected and corrected. It includes key indicators linked to the validation rules, namely cesarean section rate and prevalence of gestational diabetes, gestational hypertension, preeclampsia, preterm birth, cesarean section rate, intrauterine growth restriction, macrosomia, anemia in the first, second, and third trimesters and postpartum anemia.5Validation patterns (validation_standards)It includes clinical and epidemiological patterns used for relationship testing during validation, where expected (reference) values are compared against the values estimated for populations exposed to a risk factor, based upon scientific evidence. For instance, gestational diabetes is two times more frequent among obese women (BMI≥30), as the prevalence for exposed populations is estimated at 9.69%, versus 4.79% in reference populations [11]. Other patterns include gestational hypertension, preeclampsia cesarean rates in women with BMI≥30; gestational diabetes, hypertension, and cesarean rate in women aged 35 or older; labor induction, macrosomia and cesarean rate in women with gestational diabetes; preterm births due to preeclampsia; preterm births due to anemia; and average birth weight in intrauterine growth restriction.6Missingness (missingness)It includes missing rates for each computed variable. Only variables with missing not at random (MNAR) values will report missing rates above 0%.7Data generation rules (generation_rules)Description of the clinical reasoning, constraints, and relationships used to generate the data8Clinical assumptions (clinical_assumptions)Summary of thresholds, expected ranges, and domain-specific assumptions applied9Limitations (limitations)Description of dataset limitations, including non-auditable variables and constraints Content of the public Excel file provided for use. The complete dataset is structured into nine sheets, including the README specification, the main dataset with synthetic pregnancy trajectories, a comprehensive data dictionary, and multiple supporting sheets describing the clinical rules, assumptions, and reasoning governing variable generation, as well as dataset limitations and validation procedures. Validation was performed by means of rule-based consistency and plausibility-oriented patterns. The overview of the complete dataset file is presented in Table 1. In addition to the complete XLSX data file, the primary dataset with the synthetic records (Sheet 2) and the data dictionary (Sheet 3) are also provided as separate Comma-Separated Values (CSV) files - available in both Portuguese and English, in alignment with the FAIR (Findable, Accessible, Interoperable, Reusable) principles’ recommendations to use standard data formats [12]. The CSV versions can be freely downloaded from the Zenodo repository using the same link as the complete dataset. Therefore, six data files are publicly available: (i) the complete dataset in English (XLSX format, “Complete Pregnant patients - Synthetic Dataset.xlsx”); (ii) the complete dataset in Portuguese (XLSX format, “Complete Pregnant patients - Synthetic Dataset Portuguese.xlsx”); (iii) the primary dataset in English, consisting solely of synthetic records (CSV format, “Pregnant patients - Synthetic Dataset.csv”); (iv) the primary dataset in Portuguese, consisting solely of synthetic records (CSV format, “Pregnant patients - Synthetic Dataset_portuguese.csv”); (v) the data dictionary in English (CSV format, “Dictionary.csv”); and (vi) the data dictionary in Portuguese (CSV format “Dicionario_pt.csv”). Additionally, an R script illustrating basic predictive modelling tasks and workflows, using the primary dataset in CSV format, can be accessed through the same repository under the name “predictive_modelling_tasks_csv.R”. A second version of the R script, named “predictive_modelling_tasks_xlsx.R”, is also provided for the complete dataset in XLSX format, further demonstrating how to load and interactively visualise all supplementary sheets not directly used in the predictive modelling tasks. Each row in the main dataset corresponds to an individual record and includes variables capturing clinical information across the full continuum of pregnancy care, including preconception (if applicable), antenatal consultations, delivery outcomes, puerperium, and postnatal follow-up. The dataset is structured in a wide format, where temporal information is encoded through variable naming conventions reflecting specific clinical timepoints (e.g., consultations and trimesters) within the pregnancy care pathway. The final dataset includes a comprehensive set of 162 variables, structured across multiple contexts and healthcare stages: •Maternal baseline characteristics: baseline demographic, clinical, and anthropometric variables collected at preconception or during early stages of antenatal care, such as maternal age, height, pre-gestational weight, and derived body mass index (BMI). Maternal baseline characteristics: baseline demographic, clinical, and anthropometric variables collected at preconception or during early stages of antenatal care, such as maternal age, height, pre-gestational weight, and derived body mass index (BMI). •Antenatal care variables specified at different pregnancy care timepoints: (i) by consultation (up to 8 consultations): it includes screening parameters and results of routine antenatal tests, such as maternal weight, potassium iodide supplementation, presence of audible fetal heartbeat and Godwin score; (ii) by trimester (diagnostic and laboratory parameters specified at the first, second and third trimesters): hemoglobin levels, hepatitis B surface antigen testing, anemia (derived from hemoglobin levels), HIV-1 and HIV-2 antibody testing, indirect Coombs testing, ultrasound gestational age, rubella and toxoplasmosis antibody testing, fetal abdominal and head circumferences, fetal morphology, biparietal diameter, systolic and diastolic blood pressures, 24-h proteinuria, fetal weight percentile estimate, estimated fetal weight and femur length; (iii) by gestational week: systolic and diastolic blood pressure at week 16 and 29. Antenatal care variables specified at different pregnancy care timepoints: (i) by consultation (up to 8 consultations): it includes screening parameters and results of routine antenatal tests, such as maternal weight, potassium iodide supplementation, presence of audible fetal heartbeat and Godwin score; (ii) by trimester (diagnostic and laboratory parameters specified at the first, second and third trimesters): hemoglobin levels, hepatitis B surface antigen testing, anemia (derived from hemoglobin levels), HIV-1 and HIV-2 antibody testing, indirect Coombs testing, ultrasound gestational age, rubella and toxoplasmosis antibody testing, fetal abdominal and head circumferences, fetal morphology, biparietal diameter, systolic and diastolic blood pressures, 24-h proteinuria, fetal weight percentile estimate, estimated fetal weight and femur length; (iii) by gestational week: systolic and diastolic blood pressure at week 16 and 29. •Pregnancy complications and conditions: variables specifying conditions such as gestational hypertension, preeclampsia, gestational diabetes, and intrauterine growth restriction. Pregnancy complications and conditions: variables specifying conditions such as gestational hypertension, preeclampsia, gestational diabetes, and intrauterine growth restriction. •Pre-birth, delivery and neonatal outcomes: gestational age at delivery, delivery mode, preterm birth, macrosomia, preeclampsia, whether labor induction was performed, presence of intrauterine growth restriction and birth weight. Pre-birth, delivery and neonatal outcomes: gestational age at delivery, delivery mode, preterm birth, macrosomia, preeclampsia, whether labor induction was performed, presence of intrauterine growth restriction and birth weight. •Postpartum follow-up variables: puerperium body temperature, puerperium anemia, puerperium hemoglobin levels, postpartum uterine involution, delivery scar healing evolution, postpartum bleeding (lochia), pelvic floor status, breastfeeding status, maternal weight and potassium iodide supplementation after 4 and 6 months following birth, presence of mental health referral, Edinburgh Postnatal Depression Scale (EPDS), and postpartum contraception use. Postpartum follow-up variables: puerperium body temperature, puerperium anemia, puerperium hemoglobin levels, postpartum uterine involution, delivery scar healing evolution, postpartum bleeding (lochia), pelvic floor status, breastfeeding status, maternal weight and potassium iodide supplementation after 4 and 6 months following birth, presence of mental health referral, Edinburgh Postnatal Depression Scale (EPDS), and postpartum contraception use. Table 2 displays the released final set of variables, specified through their English name and grouped according to the pregnancy care timepoints they are typically collected, as well as respective data types. Table 2List of variables included in the synthetic dataset. Variables are organized by clinical context (maternal baseline characteristics, antenatal care, pre-birth, delivery and neonatal outcomes, and postpartum follow-up variables) and by data type (binary, categorical, continuous, ordinal, and date), as defined in the data dictionary.Table 2 dummy alt textMaternal baseline characteristicsAntenatal carePre-birth, delivery and neonatal outcomesPostpartum follow-up variablesBinaryAnemia_first_trimesterepilepsy_on_anticonvulsants,family_history_of_NT,folate_metabolism_disorder,folic_acid_started,folic_acid_stopped,HBsAg_first_trimester,high_risk_folic_acid_indication,HIV1_HIV2_antibodies_first_trimester,indirect_coombs_first_trimester,celiac_disease,inflammatory_bowel_disease,intimate_partner_violence,methotrexate_use,other_psychoactive_substance_use,potassium_iodide_started,preconception_consultation_recorded,pregestational_diabetes,previous_child_with_NTD,previous_hepatitis_B_vaccination,RhD,rubella_first_trimester,tobacco_use,alcohol_use,toxoplasmosis_first_trimester,trimethoprim_use,dating_confirmed_by_first_trimester_ultrasoundanemia_second_trimester,anemia_third_trimester,AUDIT_result,fetal_heartbeat_present_visit2,fetal_heartbeat_present_visit3,fetal_heartbeat_present_visit4,fetal_heartbeat_present_visit5,fetal_heartbeat_present_visit6,fetal_heartbeat_present_visit7,GBS_result,gestational_hypertension,HBsAg_third_trimester,HIV1_HIV2_antibodies_third_trimester,indirect_coombs_second_trimester,instructed_on_fetal_movement_counting,intrapartum_antibiotic_prophylaxis,OGTT_performed,potassium_iodide_supplementation_visit2,potassium_iodide_supplementation_visit3,potassium_iodide_supplementation_visit4,potassium_iodide_supplementation_visit5,potassium_iodide_supplementation_visit6,potassium_iodide_supplementation_visit7,referral_hospital_consultation,referred_to_gestational_diabetes_clinic,rubella_second_trimester,toxoplasmosis_second_trimester,toxoplasmosis_third_trimesterfetal_growth_restriction,gestational_age_at_delivery_weeks,labor_induction,macrosomia,preeclampsia,preterm_birthcontraceptive_method_in_use,EPDS_ge_12,lochia_normal,postpartum_anemia,potassium_iodide_supplementation_4_months,potassium_iodide_supplementation_6_months,potassium_iodide_supplementation_visit8,referred_to_mental_health,scar_healing_normalCategoricalpatient_idblood_group,cervical_cytology_status,combined_screening_riskfetal_presentation,Goodwin_risk_visit3,Goodwin_risk_visit4,Goodwin_risk_visit5,Goodwin_risk_visit6,second_trimester_anatomy_scandelivery_typebreastfeeding_4_months,breastfeeding_6_months,breastfeeding_postpartum,contraceptive_method,Goodwin_risk_visit8,pelvic_floor_statusContinuousbeta_hCG,blood_pressure_week6_diastolic,blood_pressure_week6_systolic,bmi,fasting_glucose,folic_acid_dose,height_m,hemoglobin_first_trimester,maternal_age,nuchal_translucency_value,PAPP_A,visit1_weight_kgabdominal_circumference_second_trimester,abdominal_circumference_third_trimester,biparietal_diameter_second_trimester,Bishop_score,blood_pressure_week16_diastolic,blood_pressure_week16_systolic,blood_pressure_week29_diastolic,blood_pressure_week29_systolic,estimated_fetal_weight_third_trimester,femur_length_second_trimester,femur_length_third_trimester,gestational_age_at_first_trimester_ultrasound_days,Goodwin_score_visit3,Goodwin_score_visit4,Goodwin_score_visit5,Goodwin_score_visit6,head_circumference_second_trimester,head_circumference_third_trimester,hemoglobin_second_trimester,hemoglobin_third_trimester,OGTT_1h_glucose,OGTT_2h_glucose,OGTT_fasting_glucose,repeated_third_trimester_blood_pressure_diastolic,repeated_third_trimester_blood_pressure_systolic,second_trimester_biometry,third_trimester_24h_proteinuria_mg,visit2_weight_kg,visit3_weight_kg,visit4_weight_kg,visit5_weight_kg,visit6_weight_kg,visit7_weight_kgbirth_weight_g,gestational_age_at_delivery_weeks, total_weight_gainEPDS_score,Goodwin_score_visit8,postpartum_blood_pressure_diastolic,postpartum_blood_pressure_systolic,postpartum_hemoglobin,postpartum_temperature,postpartum_weight_loss,uterine_fundal_height_postpartum,visit8_weight_kg,weight_4_months_postpartum_kg,weight_6_months_postpartum_kgOrdinalestimated_fetal_weight_third_trimester_percentileDateestimated_due_date,first_trimester_ultrasound_date,last_menstrual_period,preconception_consultation_date,visit1_dateultrasound_adjusted_estimated_due_date,visit2_date,visit3_date,visit4_date,visit5_date,visit6_date,visit7_datedelivery_datedate_4_months_postpartum,date_6_months_postpartum,visit8_date List of variables included in the synthetic dataset. Variables are organized by clinical context (maternal baseline characteristics, antenatal care, pre-birth, delivery and neonatal outcomes, and postpartum follow-up variables) and by data type (binary, categorical, continuous, ordinal, and date), as defined in the data dictionary. Fig. 1 further provides a schematic representation of the dataset’s longitudinal structure, illustrating the clinical domains and temporal organization of variables across the several pregnancy care timepoints, while also highlighting key variables for which longitudinal trajectories can be reconstructed from the dataset. Schematic representation of the synthetic dataset’s longitudinal structure. The dataset contains both base variables (e.g., laboratory measurements) and derived variables (e.g., diagnostic flags), which are computed according to explicit derivation rules detailed in Sheet 7 (Table 1). Binary variables are encoded using a 0/1 scheme, while categorical variables adhere to restricted value sets to ensure internal consistency. Date type variables are formatted according to the “yyyy-mm-dd" standard. Temporal consistency is ensured through variables specified at clinically appropriate stages. Trimester-specific variables are consistent with their corresponding gestational periods, while pregnancy outcome variables are restricted to delivery and postpartum stages. Moreover, repeated measurements (e.g., maternal weight per trimester or breastfeeding status) collected throughout antenatal care enable reconstruction of individual patient trajectories over time. Additionally, the dataset incorporates realistic variability through controlled distributions, including structured missing values in non-critical variables (missing not at random), variations in clinical measurements, and differences in adherence-related variables. Furthermore, clinical and temporal constraints were applied to enhance the coherence of individual trajectories. Overall distributions regarding pregnancy outcomes, including pregnancy complications such as gestational diabetes and hypertension, as well as demographics, clinical, and anthropometric variables, were constrained to be clinically plausible and compatible with low-risk populations (Fig. 2). In obstetrics, low-risk pregnancy is the one associated with no identified risk factors for the mother or in the fetus during antenatal care [13]. However, the model was induced to include records reporting gestational diabetes and hypertension, as these conditions may appear in low-risk populations at lower rates [14,15]. Examples of established and clinically plausible benchmark values for outcome prediction and simulation tasks supported by the dataset include: preterm birth rates ranging between 4 to 16% [16]; cesarean delivery rates aligned with the recommended 10-15% benchmark [17]; share of pregnant women aged 35 or older is approximately 20% [18]; prevalence rates of approximately 5% for gestational diabetes [19] and between 5% and 10% for gestational hypertension [20]; preeclampsia complicating approximately 2 to 8% of all pregnancies worldwide [21]; and macrosomia prevalence estimated at approximately 12% of normal pregnancies globally, although reported rates vary substantially across countries, ranging between 1.6% to 28% of all births [22]. Distribution of maternal characteristics and obstetric outcomes in the synthetic dataset. Similarly, to improve statistical coherence, selected correlations and dependency patterns among variables were explicitly incorporated into the synthetic population [e.g., systolic (SBP) and diastolic (DBP) blood pressure measurements at 29 weeks are moderately correlated with gestational hypertension] (Fig. 3). Additionally, BMI (Body Mass Index) and OGTT (Oral Glucose Tolerance Test) were included as relevant clinical variables in the analysis. Key correlations in the synthetic obstetric dataset. Furthermore, established clinical associations between pregnancy outcomes and major maternal risk factors can be evaluated when using the dataset in regression analysis. Examples of epidemiological benchmark patterns that can be reproduced include: an approximately 2-fold increased risk of gestational diabetes amongst obese women [23] and amongst women aged 35 years or older compared with those younger than 25 years [24]; an approximately 1.6- to 1.8-fold increased risk of gestational hypertension and a 1.6- to 1.7-fold increased risk of preeclampsia for every 5-unit increase in BMI [25]; 2- to 3-fold higher odds of macrosomia amongst obese women [26]; and a 4- to 5-fold increased risk of preterm delivery amongst women who develop preeclampsia [27]. The dataset is fully synthetic and does not contain any real patient information. It is intended to serve educational purposes and support methodological research in synthetic data generation and obstetrics, machine learning models development, simulation and other exploratory analyses. FAIR principles were adopted to ensure that data are findable by means of standardized metadata, accessible via open resources, interoperable through shared formats (e.g., CSV) and clinical vocabularies, and reusable for both humans and machines [12]. Experimental Design, Materials and Methods Synthetic data generation was performed using the ChatGPT web interface. During the generation process, the interface displayed the model identifier “GPT-5.3”, which is therefore recorded and reported in this study. Generation was conducted under default platform settings without API access, explicit hyperparameter specification, seed control, temperature adjustment, top-p configuration, or programmatic implementation. ChatGPT’s promising capabilities to produce and handle synthetic medical records at the patient level has been described elsewhere in the literature across distinct clinical domains [7,28,29]. The generation process was conducted using a structured prompt engineering approach, in which prompts were iteratively refined across four stages (Prompt 1–4) based on feedback from health and data science experts. Full original prompts are available in the Supplementary Material 1 (in Portuguese). At each iteration, experts evaluated data quality by checking for overall errors and providing guidance across key dimensions, namely: (i)completeness: ensure that all included generated variables capture both established mandatory and non-mandatory clinical parameters required to represent complete low-risk pregnancy care pathways, from preconception to post birth follow up; completeness: ensure that all included generated variables capture both established mandatory and non-mandatory clinical parameters required to represent complete low-risk pregnancy care pathways, from preconception to post birth follow up; (ii)auditability: experts assessed the transparency of data generation and validation, and verified whether the data dictionary and metadata provide sufficient information for adequate audit, interpretation and reproducibility; auditability: experts assessed the transparency of data generation and validation, and verified whether the data dictionary and metadata provide sufficient information for adequate audit, interpretation and reproducibility; (iii)semantic accuracy and reliability: experts checked the semantic accuracy and reliability of individual variable by assessing how close data values are to a set of domain-specific values considered semantically correct, and the degree to which generated values are regarded as authentic and believable by potential users; semantic accuracy and reliability: experts checked the semantic accuracy and reliability of individual variable by assessing how close data values are to a set of domain-specific values considered semantically correct, and the degree to which generated values are regarded as authentic and believable by potential users; (iv)consistency: experts ensured internal coherence within each record (e.g., hemoglobin levels in the third trimester should be <10.5 g/dL when the variable computing anemia diagnosis is flagged as present in that period): consistency: experts ensured internal coherence within each record (e.g., hemoglobin levels in the third trimester should be <10.5 g/dL when the variable computing anemia diagnosis is flagged as present in that period): (v)statistical coherence: data distributions were examined to check whether they are clinically plausible and compatible with low-risk populations. statistical coherence: data distributions were examined to check whether they are clinically plausible and compatible with low-risk populations. Initial prompt design (Prompt 1) The model was initially prompted with a list of essential variables, defined according to the National Low-Risk Pregnancy Surveillance Program [30] from the Portuguese National Health Service. Instructions required the LLM to act as a domain expert, and encode rules for low-risk pregnancies, such as visits and event patterns, plausible clinical parameters and outcomes ranges, and compliance with mandatory known relationships between variables. No real ground truth or patient data were used, thereby limiting the data generation process to a zero-shot setting. Inclusion criteria constrained data generation to low-risk profiles. ChatGPT output was designed to include the main dataset in tabular format, either as CSV or Excel (XLSX) file including a self-explainable data dictionary, rules governing data generation, clinical assumptions considered, validation report and data limitations. Key specified elements included: (i)Definition of each row as an individual pregnant woman Definition of each row as an individual pregnant woman (ii)Longitudinal structure from preconception/first consultation to delivery Longitudinal structure from preconception/first consultation to delivery (iii)Specification of temporal collection points (e.g., gestational weeks) Specification of temporal collection points (e.g., gestational weeks) (iv)Enforcement of general clinical constraints: Enforcement of general clinical constraints: (v)causal relationships (e.g., BMI derived from weight and height) causal relationships (e.g., BMI derived from weight and height) (vi)temporal coherence of events temporal coherence of events (vii)Requirement for realistic statistical distributions and epidemiological correlations Requirement for realistic statistical distributions and epidemiological correlations (viii)Inclusion of variability to avoid deterministic or “perfect” profiles Inclusion of variability to avoid deterministic or “perfect” profiles (ix)Output requirements including dataset, data dictionary, and generation rules Output requirements including dataset, data dictionary, and generation rules This initial version focused on high-level clinical plausibility and structure, without enforcing strict derivation rules or auditability constraints. Prompt refinement following first expert feedback (Prompt 2) Following a first expert review for general errors and inconsistencies, the prompt was refined to incorporate expert feedback and include explicit derivation rules, stricter clinical reasoning, and improve pregnancy trajectory variability. Main prompt modifications can be summarized as follows: (i)Explicit derivation of variables: ChatGPT was required to add binary flag variables to be systematically derived from underlying variables, e.g., EPDS flag indicating possible depression equals 1 when the underlying EPDS score is 12 or higher, Goodwin score categorization (0–2 low; 3–6 medium; ≥7 high), standard folic acid intake equal to 400 µg, or 5000 µg only if BMI ≥ 30. Explicit derivation of variables: ChatGPT was required to add binary flag variables to be systematically derived from underlying variables, e.g., EPDS flag indicating possible depression equals 1 when the underlying EPDS score is 12 or higher, Goodwin score categorization (0–2 low; 3–6 medium; ≥7 high), standard folic acid intake equal to 400 µg, or 5000 µg only if BMI ≥ 30. (ii)Enhanced clinical reasoning: mutual exclusivity rules were introduced at this stage to improve clinical plausibility and coherence, e.g. indication of pre-existing diabetes excluding gestational diabetes. Detailed epidemiological associations were explicitly included by considering the relationship between the exposure to certain risks with specific conditions or events, namely: High BMI associated with higher risk of gestational diabetes and hypertension; Advanced maternal age associated with higher risk of gestational diabetes, hypertension, premature birth, cesarean delivery, and chromosomal abnormalities; Gestational diabetes associated with higher risk of labor induction or cesarean delivery; Preeclampsia associated with higher risk of premature birth and cesarean delivery; Intrauterine growth restriction associated with higher risk of low birth weight. Enhanced clinical reasoning: mutual exclusivity rules were introduced at this stage to improve clinical plausibility and coherence, e.g. indication of pre-existing diabetes excluding gestational diabetes. Detailed epidemiological associations were explicitly included by considering the relationship between the exposure to certain risks with specific conditions or events, namely: High BMI associated with higher risk of gestational diabetes and hypertension; Advanced maternal age associated with higher risk of gestational diabetes, hypertension, premature birth, cesarean delivery, and chromosomal abnormalities; Gestational diabetes associated with higher risk of labor induction or cesarean delivery; Preeclampsia associated with higher risk of premature birth and cesarean delivery; Intrauterine growth restriction associated with higher risk of low birth weight. (iii)Controlled variability and realism to prevent the generation of templated data: new rules were introduced to enhance variability and realism within the synthetic population. These include the incorporation of a missingness rate (MR) ranging from 3 to a maximum of 7% in non-critical variables (missing not at random), in particular non-mandatory screening tests such as oral glucose tolerance test (OGTT) and supplement prescription (e.g. folic acid). Preconception visit date is also constrained to have missing values, considering that the share of women receiving preconception care remains inadequate and about 40% of pregnancies do not receive the necessary services before pregnancy, according to the World Health Organization [31]. Variables with missing values thereby included preconception visit date (MR=72.5%), OGTT results at fasting and at 1 h and 2 h following glucose solution intake (MR=1.9%), cervical cytology status (MR=5.1%), pelvic floor status (MR=4%), folic acid dosage (MR= 8.1%), and AUDIT-C test results (MR=6.7%). Further variability was modelled by including an estimated 5 to 10% rate of non-adherence to potassium iodide supplementation, as well as minor administrative inconsistencies that do not compromise clinical coherence (e.g., cervical cytology status may occasionally be left blank for patients who have not undergone the test, despite the admission of a “Not performed” category). Controlled variability and realism to prevent the generation of templated data: new rules were introduced to enhance variability and realism within the synthetic population. These include the incorporation of a missingness rate (MR) ranging from 3 to a maximum of 7% in non-critical variables (missing not at random), in particular non-mandatory screening tests such as oral glucose tolerance test (OGTT) and supplement prescription (e.g. folic acid). Preconception visit date is also constrained to have missing values, considering that the share of women receiving preconception care remains inadequate and about 40% of pregnancies do not receive the necessary services before pregnancy, according to the World Health Organization [31]. Variables with missing values thereby included preconception visit date (MR=72.5%), OGTT results at fasting and at 1 h and 2 h following glucose solution intake (MR=1.9%), cervical cytology status (MR=5.1%), pelvic floor status (MR=4%), folic acid dosage (MR= 8.1%), and AUDIT-C test results (MR=6.7%). Further variability was modelled by including an estimated 5 to 10% rate of non-adherence to potassium iodide supplementation, as well as minor administrative inconsistencies that do not compromise clinical coherence (e.g., cervical cytology status may occasionally be left blank for patients who have not undergone the test, despite the admission of a “Not performed” category). (iv)Variable-level constraints: several variable constraints were introduced, such as enforcement of integer-only scores whenever applicable (e.g., Goodwin score), properly structured laboratory measurements (e.g., 2-h Oral Glucose Tolerance Test taken at 0, 1, and 2-h, with automated derivation of gestational diabetes when any one value meets or exceeds the clinical threshold). Variable-level constraints: several variable constraints were introduced, such as enforcement of integer-only scores whenever applicable (e.g., Goodwin score), properly structured laboratory measurements (e.g., 2-h Oral Glucose Tolerance Test taken at 0, 1, and 2-h, with automated derivation of gestational diabetes when any one value meets or exceeds the clinical threshold). The resulting dataset following prompt refinement transitioned from general plausibility to a more rule-driven synthetic data generation with the inclusion of more domain-specific knowledge, as well as controlled noise, variability, and realism. Prompt refinement focusing on clinical modelling (Prompt 3) The third prompt was further refined to specify full clinical pathways and add quantitative constraints, focusing on formal clinical modelling aspects. Main additions included: (i)Explicit longitudinal modelling: full pregnancy trajectory was established, namely by means of consultations (1 to 8 antenatal consultations), trimesters, delivery, puerperium, and follow-up variables. Explicit longitudinal modelling: full pregnancy trajectory was established, namely by means of consultations (1 to 8 antenatal consultations), trimesters, delivery, puerperium, and follow-up variables. (ii)Quantitative constraints: detailed modelling of realistic gestational weight gain by BMI category, postpartum weight loss ranges, and physiological variability across consultations. Quantitative constraints: detailed modelling of realistic gestational weight gain by BMI category, postpartum weight loss ranges, and physiological variability across consultations. (iii)Temporal logic enforcement: inclusion of strict rules ensuring that certain events only occur within valid gestational time frames, and that dynamic variables will evolve over time (e.g., maternal weight). Temporal logic enforcement: inclusion of strict rules ensuring that certain events only occur within valid gestational time frames, and that dynamic variables will evolve over time (e.g., maternal weight). (iv)Epidemiological dependency modelling: explicit encoding of epidemiological associations identified in the scientific literature was specified, namely the relationship between BMI with gestational diabetes and hypertension, preeclampsia with preterm birth, and Intrauterine growth restriction (IUGR) with low birth weight. Epidemiological dependency modelling: explicit encoding of epidemiological associations identified in the scientific literature was specified, namely the relationship between BMI with gestational diabetes and hypertension, preeclampsia with preterm birth, and Intrauterine growth restriction (IUGR) with low birth weight. This stage established a stricter synthetic data generation, adding stochastic variability and ensuring reproducible clinical reasoning. Prompt refinement focusing on auditability and schema standardization (Prompt 4) Following the last expert panel review, feedback pointed to the need for a final refinement focusing on data auditability, improved schema rigor, and validation transparency, aligning the dataset with essential data quality and reproducibility requirements. Key enhancements included: (i)Auditability requirements: all derived variables must specify explicit base variables (for derived variables), documented derivation rules and clinical reasoning governing variable generation, and non-auditable variables must be explicitly flagged and justified. Auditability requirements: all derived variables must specify explicit base variables (for derived variables), documented derivation rules and clinical reasoning governing variable generation, and non-auditable variables must be explicitly flagged and justified. (ii)Standardized data schema: definition of a mandatory metadata for each variable specifying, semantic type, exported type, coding scheme definitions, allowed values and derivation logic. Standardized data schema: definition of a mandatory metadata for each variable specifying, semantic type, exported type, coding scheme definitions, allowed values and derivation logic. (iii)Encoding constraints: binary variables were enforced as 0/1 only, with no text labels, whereas categorical variables required closed value sets. Encoding constraints: binary variables were enforced as 0/1 only, with no text labels, whereas categorical variables required closed value sets. (iv)Hard validation rules: quantitative constraints enforced weight gain ranges, postpartum weight loss, and glucose testing consistency. Hard validation rules: quantitative constraints enforced weight gain ranges, postpartum weight loss, and glucose testing consistency. (v)Pre-export validation: mandatory row-level validation including temporal consistency, cross-verification of base and derived variables, and absence of rule violations. The model was requested to regenerate the dataset whenever non-compliant records were detected. Pre-export validation: mandatory row-level validation including temporal consistency, cross-verification of base and derived variables, and absence of rule violations. The model was requested to regenerate the dataset whenever non-compliant records were detected. (vi)Transparency requirements: Mandatory outputs were specified and must include a validation report, auditability documentation, and explicit dataset limitations. Transparency requirements: Mandatory outputs were specified and must include a validation report, auditability documentation, and explicit dataset limitations. This final prompt refinement stage thereby focused on ensuring traceability, auditability, reproducibility, and compliance with structured data quality principles. Specific generation rules to support auditing are presented in Table 3. where each derived variable from the dataset is explicitly linked to its corresponding base variables through the names provided in the primary variable list (Table 2), thereby supporting the traceability, auditability, and transparency of the synthetic data generation process. Derivation logic, as well as coherence constraints are also specified whenever applicable. Variables not listed in Table 3 followed overall synthetic generation instructions specified via prompting, thereby not presenting base variables for auditing. Additionally, clinical assumptions and statistical patterns considered for dataset evaluation and validation are displayed in Tables 4 and 5, respectively. It is important to reinforce that generation rules, as well as validation patterns and clinical patterns constraints were designed to maximize auditability, internal consistency, and clinical plausibility, rather than reproduce a specific real-world distribution. Table 3Generation rules.Table 3 dummy alt textDerivate variablesDerivation logicBase VariablesGeneration rulesBody Mass Index (bmi)BMI derived from the baseline weight and height recorded on the first visit.visit1_weight_kg, height_mbmi = visit1_weight_kg/(height_m)2Preconception consultation date (preconception_consultation_date)Date of preconception consultation, whenever applicable. It must be prior to the last menstrual period.preconception_consultation_recorded,last_menstrual_periodifpreconception_consultation_recorded = 1, then preconception_consultation_date filled with synthetic date, missing otherwise.Constraint:preconception_consultation_date < last_menstrual_period.Contraception use (contraceptive_method_in_use)Contraception use is flagged during postpartum follow-up period whenever informed by the pregnantcontraceptive_methodIfcontraceptive_method ≠ “none”, then contraceptive_method_in_use = 1, 0 otherwise.Gestational age at first trimester ultrasound (gestational_age_at_first_trimester_ultrasound_days)Gestational age in days at the first trimester ultrasound is computed as the difference between the ultrasound date and the last period date, adjusted for synthetic ultrasound discrepancy.first_trimester_ultrasound_date,last_menstrual_periodgestational_age_at_first_trimester_ultrasound_days= (first_trimester_ultrasound_date - last_menstrual_period) + synthetic discrepancy ranging between ± 10 daysEstimated due date (estimated_due_date, ultrasound_adjusted_estimated_due_date)Estimated due date is computed as 280 days after the last period date.An optional adjustment is available by considering the definitive due date estimated by the first trimester ultrasound.last_menstrual_period,first_trimester_ultrasound_date, gestational_age_at_first_trimester_ultrasound_daysestimated_due_date = last_menstrual_period + 280if [gestational_age_at_first_trimester_ultrasound_days – (first_trimester_ultrasound_date - last_menstrual_period)] ≥ 7 then:ultrasound_adjusted_estimated_due_date = first_trimester_ultrasound_date + (280 - gestational_age_at_first_trimester_ultrasound_days)elseultrasound_adjusted_estimated_due_date = estimated_due_dateDelivery date (delivery_date)Delivery date is the last menstrual period (date) plus rounding of gestational age in weeks at delivery.last_menstrual_period,gestational_age_at_delivery_weeksdelivery_date = last_menstrual_period + (gestational_age_at_delivery_weeks*7)Postpartum follow-up consultation dates (visit8_date, date_4_months_postpartum,date_6_months_postpartum)Date of the postpartum check-up (visit 8), usually performed between 28 to 42 days after birth, as well as follow-up dates reported approximately at 4 and 6 months postpartum.delivery_datevisit8_date = delivery_date + [28-42].date_4_months_postpartum = delivery_date +≈ 4 months.date_6_months_postpartum = delivery_date +≈ 6 months.Gestational age at delivery (gestational_age_at_delivery_weeks)Gestational age at delivery in weeks is computed as the difference between delivery date and reference gestational age (confirmed by third trimester ultrasound).delivery_date, ultrasound_adjusted_estimated_due_dategestational_age_at_delivery_weeks ≈ 40 + ((delivery_date - ultrasound_adjusted_estimated_due_date)/7).Constraint:Decimal value ranging between 20 to 43 weeks. Majority between 37 and 41 weeks.Gestational weight gain (total_weight_gain)Total gestational weight gain considering weight measurements across the entire antenatal period (from visit 1 to visit 7), further constrained according to baseline BMI-specific clinical ranges.visit1_weight_kg, visit7_weight_kg, bmitotal_weight_gain = visit7_weight_kg - visit1_weight_kg*Constraint:if bmi < 18.5, then total_weight_gain ≥ 12.5 AND total_weight_gain < 18if bmi ≥ 18.5 and ≤ 24.9, then total_weight_gain ≥ 11.5 and total_weight_gain ≤ 16if bmi ≥ 25 and ≤ 29.9, then total_weight_gain ≥ 7 and total_weight_gain ≤ 11.5if bmi ≥ 30, then total_weight_gain ≥ 5 and total_weight_gain ≤ 9* Institute of Medicine and National Research Council (IOM and NRC) recommendations [32].Postpartum weight loss (postpartum_weight_loss)Postpartum weight loss (postpartum_weight_loss) is defined as the difference between the weight in the last antenatal consultation (visit7_weight_kg) and the weight recorded during the postpartum period (visit8_weight_kg). Postpartum weight loss is constrained between 5 and 8 kg.visit7_weight_kg, visit8_weight_kg, postpartum_weight_losspostpartum_weight_loss = visit7_weight_kg – visit8_weight_kgConstraints:postpartum_weight_loss ≥ 5 AND postpartum_weight_loss ≤8Anemia derivation (anemia_first_trimester, anemia_second_trimester, anemia_third_trimester, postpartum_anemia)Anemia flags derived from hemoglobin values using trimester-specific clinical thresholds (hemoglobin_first_trimester… hemoglobin_third_trimester) and postpartum recording (postpartum_anemia).hemoglobin_first_trimester, hemoglobin_second_trimester, hemoglobin_third_trimester, postpartum_hemoglobinIf hemoglobin_first_trimester < 11 g/dL then anemia_first_trimester=1, 0 otherwise.If hemoglobin_second_trimester < 10.5 g/dL then anemia_second_trimester=1, 0 otherwise.If hemoglobin_third_trimester < 10.5 g/dL then anemia_third_trimester=1, 0 otherwise.If postpartum_hemoglobin < 10 g/dL then postpartum_anemia =1, 0 otherwise.Oral Glucose Tolerance Test (OGTT_performed)OGTT performed only if fasting glucose <92 mg/dL (fasting_glucose), otherwise not performed. If performed, additional recordings are computed after 1 and 2 h following glucose solution intake.fasting_glucose,OGTT_fasting_glucose, OGTT_1h_glucose, OGTT_2h_glucoseif fasting_glucose < 92 mg/dL then OGTT_performed =1, 0 otherwise.Goodwin risk (Goodwin_risk_visit3, Goodwin_risk_visit4, Goodwin_risk_visit5,Goodwin_risk_visit6,Goodwin_risk_visit8)Categorical Goodwin risk thresholds are derived from Goodwin scores evaluated longitudinally across weeks 20 to 36 (Goodwin_score_visit3… Goodwin_score_visit6) and during postpartum period (Goodwin_score_visit8).Goodwin_score_visit3,Goodwin_score_visit4,Goodwin_score_visit5,Goodwin_score_visit6,Goodwin_score_visit8If [Goodwin_score_visit3 ≥ 0 and Goodwin_score_visit3 ≤ 2], then Goodwin_risk_visit3 = “low” else if [Goodwin_score_visit3 ≥ 3 and Goodwin_score_visit3 ≤ 6], then Goodwin_risk_visit3 = “medium”, else Goodwin_risk_visit3 = “high”If [Goodwin_score_visit4 ≥ 0 and Goodwin_score_visit4 ≤ 2], then Goodwin_risk_visit4 = “low” else if [Goodwin_score_visit4 ≥ 3 and Goodwin_score_visit4 ≤ 6], then Goodwin_risk_visit4 = “medium”, else Goodwin_risk_visit4 = “high”If [Goodwin_score_visit5 ≥ 0 and Goodwin_score_visit5 ≤ 2], then Goodwin_risk_visit5 = “low” else if [Goodwin_score_visit5 ≥ 3 and Goodwin_score_visit5 ≤ 6], then Goodwin_risk_visit5 = “medium”, else Goodwin_risk_visit5 = “high”If [Goodwin_score_visit6 ≥ 0 and Goodwin_score_visit6 ≤ 2], then Goodwin_risk_visit6 = “low” else if [Goodwin_score_visit6 ≥ 3 and Goodwin_score_visit6 ≤ 6], then Goodwin_risk_visit6 = “medium”, else Goodwin_risk_visit6 = “high”If [Goodwin_score_visit8 ≥ 0 and Goodwin_score_visit8 ≤ 2], then Goodwin_risk_visit8 = “low” else if [Goodwin_score_visit8 ≥ 3 and Goodwin_score_visit8 ≤ 6], then Goodwin_risk_visit8 = “medium”, else Goodwin_risk_visit8 = “high”Edinburgh Postnatal Depression Scale (EPDS) derivation (EPDS_ge_12)EPDS flag indicating whether the computed EPDS score is 12 or higher is derived from the continuous EPDS score recorded during postpartum period (EPDS_score).EPDS_scoreIf EPDS_score ≥ 12 then EPDS_ge_12 = 1, 0 otherwise.Folic acid dosage (folic_acid_dose) and indication of high-risk folic acid intake (high_risk_folic_acid_indication)Folic acid dosage is constrained to a standard dose of 400 µg; dosages corresponding to 5000 µg are restricted to high-risk clinical conditions. Indication on whether acid folic intake is inappropriate are derived from the presence of any high-risk clinical condition.folic_acid_started, bmi, pregestational_diabetes, previous_child_with_NTD, family_history_of_NT, epilepsy_on_anticonvulsants, folate_metabolism_disorder, celiac_disease, inflammatory_bowel_disease, methotrexate_use, trimethoprim_use.if bmi ≥ 30 or if [previous_child_with_NTD or family_history_of_NT or pregestational_diabetes or epilepsy_on_anticonvulsants or folate_metabolism_disorder or celiac_disease or inflammatory_bowel_disease or methotrexate_use or trimethoprim_use] = 1, then high_risk_folic_acid_indication = 1, otherwise 0.If folic_acid_started = 0 thenfolic_acid_dose = “NA”, else if high_risk_folic_acid_indication = 1 then folic_acid_dose =5000, else folic_acid_dose =400.Potassium iodide supplementation(potassium_iodide_started)Supplementation with potassium iodide is initiated as 1 and may be mantained throughout pregnancy and postpartum follow-up, allowing a 5–10% simulated non-adherence.potassium_iodide_supplementation_visit2potassium_iodide_supplementation_visit3potassium_iodide_supplementation_visit4potassium_iodide_supplementation_visit5potassium_iodide_supplementation_visit6potassium_iodide_supplementation_visit7potassium_iodide_supplementation_visit8potassium_iodide_supplementation_4_monthspotassium_iodide_supplementation_6_monthsIf potassium_iodide_started = 1, then supplementation may be recorded from visit 2 to 6 months postpartum, with 5-10% simulated non-adherence.Fetal cardiac activity (fetal_heartbeat_present_visit2,fetal_heartbeat_present_visit3,fetal_heartbeat_present_visit4,fetal_heartbeat_present_visit5,fetal_heartbeat_present_visit6,fetal_heartbeat_present_visit7)Audible fetal heartbeat expected after fetal viability and maintained in subsequent consultations within antenatal care (visits 2 to 7).fetal_heartbeat_present_visit2fetal_heartbeat_present_visit3fetal_heartbeat_present_visit4fetal_heartbeat_present_visit5fetal_heartbeat_present_visit6fetal_heartbeat_present_visit7Once fetal heartbeat is recorded as present, persistence is expected in subsequent antenatal visits (visits 2 to 7).Breastfeeding (breastfeeding_postpartum,breastfeeding_4_months,breastfeeding_6_months)Breastfeeding information is reported using 3 possible categories (“exclusive”, “mixed” and “formula”) and is recorded at visit 8 and at 4 and 6 months following birth, only enabling valid transitions.breastfeeding_postpartumbreastfeeding_4_monthsbreastfeeding_6_monthsAllowed longitudinal transitions are restricted to:•exclusive to exclusive, mixed or formula•mixed to mixed or formula•formula to formulaGestational Diabetes Referral (referred_to_gestational_diabetes_clinic)Gestational diabetes (pregestational_diabetes) is a synthetic baseline maternal characteristic. Referral to diabetes care may be flagged according to fasting blood glucose and/or OGTT results.OGTT_fasting_glucose, OGTT_1h_glucose, OGTT_2h_glucose,If fasting_glucose ≥ 92 mg/dL or OGTT_1h_glucose ≥ 180 mg/dL orOGTT_2h_glucose ≥ 153 md/dL then referred_to_gestational_diabetes_clinic = 1, 0 otherwise.Gestational Hypertension (gestational_hypertension)Gestational hypertension is reported from elevated systolic or diastolic blood pressure measurements taken after 20 weeks without significant proteinuria (24-h proteinuria levels below 300 mg).blood_pressure_week29_diastolic,blood_pressure_week29_systolic,repeated_third_trimester_blood_pressure_diastolic,repeated_third_trimester_blood_pressure_systolic,third_trimester_24h_proteinuria_mgIf [blood_pressure_week29_diastolic or repeated_third_trimester_blood_pressure_diastolic] ≥ 90 and if [blood_pressure_week29_systolic or repeated_third_trimester_blood_pressure_systolic] > 140 and third_trimester_24h_proteinuria_mg < 300 then gestational_hypertension = 1, 0 otherwise.Diagnosis of preeclampsia (preeclampsia)Preeclampsia is flagged from a 24-h proteinuria result ≥ 300 mg combined with elevated systolic and diastolic blood pressure measurements taken after 20 weeks.blood_pressure_week29_diastolic,blood_pressure_week29_systolic,repeated_third_trimester_blood_pressure_diastolic,repeated_third_trimester_blood_pressure_systolic,third_trimester_24h_proteinuria_mgIf [blood_pressure_week29_diastolic or repeated_third_trimester_blood_pressure_diastolic] ≥ 90 and if [blood_pressure_week29_systolic or repeated_third_trimester_blood_pressure_systolic] > 140 and third_trimester_24h_proteinuria_mg ≥ 300 then preeclampsia = 1, 0 otherwise.Intrauterine growth restriction (IUGR) (fetal_growth_restriction)IUGR is flagged when the synthetic estimated fetal weight percentile in the third trimester is lower than 10.estimated_fetal_weight_third_trimester_percentileIf estimated_fetal_weight_third_trimester_percentile < 10 then fetal_growth_restriction = 1, 0 otherwise.Macrosomia (macrosomia)Macrosomia is flagged when birth weight is 4000g or higher.birth_weight_gIf birth_weight_g ≥ 4000 then macrosomia = 1, 0 otherwise.Preterm birth (preterm_birth)Preterm birth is flagged when gestational age at delivery is lower than 37 weeks.gestational_age_at_delivery_weeksIf gestational_age_at_delivery_weeks < 37 then preterm_birth = 1, 0 otherwise. Generation rules. •exclusive to exclusive, mixed or formula exclusive to exclusive, mixed or formula •mixed to mixed or formula mixed to mixed or formula •formula to formula formula to formula Table 4Clinical assumptions.Table 4 dummy alt textPatternDescriptionBMI and complicationsHigher BMI associated with increased probability of gestational diabetes, hypertension and cesarean deliveryAdvanced maternal ageMaternal age ≥35 is associated with higher risk of diabetes, hypertension, prematurity and cesarean deliveryGestational diabetesAssociated with increased probability of labor induction, macrosomia and cesarean deliveryPreeclampsiaAssociated with increased probability of preterm birthIUGRAssociated with lower birth weightAnemiaAssociated with increased probability of preterm birth Clinical assumptions. Table 5Statistical patterns.Table 5 dummy alt textAssumptionDescriptionTemporal consistencyVariables are constrained to clinically appropriate gestational timepointsLongitudinal structureEach record represents a full trajectory from preconception to postpartum follow-upBiological plausibilityClinical measurements follow realistic physiological distributionsLaboratory testing logicScreening tests follow standard obstetric care guidelinesBreastfeeding trajectoriesAllowed transitions: exclusive, mixed, artificial; reverse transitions not permittedMissingness3–7% structured missing values introduced in non-critical variablesVariabilityControlled stochastic variability introduced to avoid deterministic patternsAuditabilityDerived variables are linked to base variables whenever available in the dataset Statistical patterns. Limitations The dataset is restricted to low-risk profiles, and it should not be used for clinical inference, institutional benchmarking, or causal estimation, neither to directly support clinical decision-making and clinical model deployment. Due to its synthetic nature, rare complications and the full demographic and clinical diversity of real populations may be underrepresented. Moreover, data should not be assumed to reproduce real epidemiological distributions within specific institutions, geographic regions or national populations. Fetal growth curves, biometrics, and percentiles are simplified synthetic approximations designed for internal consistency and should not be interpreted as substitutes for established clinical standards. Preeclampsia modelling is simplified, excluding the full spectrum of laboratory criteria and target-organ involvement in favor of auditability over clinical detail. Social factors are not comprehensively represented, with key psychosocial variables such as social support and education omitted. The etiology underlying positive second-trimester Coombs test results is not explicitly modelled. Positive cases were kept as rare and internally consistent; however, variables that could reveal the underlying causes of positivity (e.g., autoimmune disease, inducing medications, prior blood transfusion) were not included, making individual etiological assessment impossible. The first-trimester combined risk assessment is based on a simplified rule using nuchal translucency, PAPP-A, β-hCG, and maternal age, and therefore does not replicate the full complexity of clinical trisomy screening algorithms. Finally, formal privacy-risk evaluation metrics were not the focus of this release, since the dataset was generated without direct use of real patient-level records. Ethics Statement This article describes a fully synthetic dataset generated without reference to real individuals. No real patient-level data were used for training, generation, evaluation, or validation. No human participants or identifiable data were involved. Therefore, ethics approval was not required. CRediT authorship contribution statement Gonçalo Pinto: Methodology, Software, Writing – original draft. Vítor Crista: Methodology, Supervision, Writing – review & editing. Regina Silva: Writing – review & editing. Bruno Costa: Validation, Writing – review & editing. Filipe Andrade Bernardi: Validation, Writing – review & editing. Rute Almeida: Validation, Writing – review & editing. Julio Souza: Writing – original draft. Alberto Freitas: Project administration, Funding acquisition, Writing – review & editing. Goreti Marreiros: Supervision, Project administration, Funding acquisition, Writing – review & editing. Luís Conceição: Methodology, Supervision, Project administration, Funding acquisition, Writing – review & editing. Acknowledgments MedGPT project (ITEA-2023-23020), funded by the European Regional Development Fund (ERDF) and Portugal 2030 through the Innovation and Digital Transition Program, under project number NORTE2030-FEDER-01414600, and the Portuguese Foundation for Science and Technology (FCT) (project n° UIDB/00760/2020). Rute Almeida acknowledges support from the Portuguese Foundation for Science and Technology (FCT) for a tenure-funded position at the Faculty of Medicine of the University of Porto (2023.13694.TENURE.056). Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Footnotes Supplementary material associated with this article can be found, in the online version, at doi:10.1016/j.dib.2026.112933. Contributor Information Gonçalo Pinto, Email: gvtpo@isep.ipp.pt. Julio Souza, Email: jbs@isep.ipp.pt. Supplementary materials Data Availability (Zenodo).Low-Risk Synthetic Patient Level Pregnancy Trajectories (Original data) (Zenodo).Low-Risk Synthetic Patient Level Pregnancy Trajectories (Original data) References Associated Data Supplementary Materials Data Availability Statement (Zenodo).Low-Risk Synthetic Patient Level Pregnancy Trajectories (Original data) (Zenodo).Low-Risk Synthetic Patient Level Pregnancy Trajectories (Original data)