Table 1 Description and characterization of included studies

Prediction of breastfeedig practices

 Oliver-Roig et al. (2022) Spain [23]

Baseline; Logistic Regression; Naive Bayes; Decision Tree; Random Forest; AdaBoost (Adaptive Boosting); XGBoost, CatBoost and LightGBM; Support Vector Machine; Neural Network (Multilayer Perceptron); Nearest Neighbors

2042 healthy primiparous mothers giving birth in eighteen hospitals in Eastern Spain

To predict exclusive breastfeeding during the postpartum hospital stay using ML algorithms and explain the behavior of the ML model to support decision-making

The results showed that the XGBoost algorithm achieved the best performance in terms of AUC-ROC and Brier score. Additionally, important variables such as pacifier use, and maternal self-efficacy were identified in predicting exclusive breastfeeding. The model demonstrated the ability to predict the probability of exclusive breastfeeding for specific cases

 Silva et al. (2021) Brazil [24]

Decision tree models adjusted using the CART (Classification and Regression Trees) algorithm

Longitudinal study of 1003 infants in referral centers in Brazil for high-risk fetuses and neonates

To build a tree-based analysis to determine the variables that can predict breastfeeding patterns at hospital discharge and at 3 and 6 months of age in high-risk infant referral centers

These models were used to predict infant feeding practices based on various predictor variables. The mean model accuracy in cross-validation was 83% at hospital discharge, 63% at 3 months, and 50% at 6 months; indicating that decision tree models were able to reasonably predict infant feeding practices at these three key postnatal periods using variables such as hospital stay duration, type of feeding received during hospitalization, and other maternal and neonatal characteristics as predictors

 Sampieri et al. (2022) Mexico [25]

Supervised machine learning methods. Attribute selection algorithms generated by supervised learning. Bayesian networks and decision trees models for classification. Algorithms implemented by Weka (Bayes Net and J48) were used to analyze classification methods

Descriptive study analyzing data from ENADID 2018

n = 26,578 mothers between 15 and 54 years

To determine the association of skin-to-skin contact between mothers and their newborns immediately after birth with the initiation of breastfeeding within the first hour of life, breastfeeding duration ≥ 6 months, and the introduction of breast milk substitutes

Skin-to-skin contact is associated with vaginal delivery, receiving an explanation about breastfeeding after delivery, initiating breastfeeding within the first hour of life, having ever breastfed, and a breastfeeding duration of at least 6 months. Additionally, breastfeeding duration in days was greater in the skin-to-skin contact group compared to the non-skin-to-skin contact group

Analyses using Bayesian networks and decision trees indicated a probabilistic dependency relationship between skin-to-skin contact and receiving an explanation about breastfeeding after delivery. In summary, AI helped identify significant relationships between the studied variables and breastfeeding

 Elgersma et al. (2023) United States [26]

Supervised learning techniques: elastic net regression, multiple data imputation, and variable importance analysis. Elastic net regression was applied for predictor variable selection and final model construction for each imputed data set

Infants with complex congenital heart disease (single ventricle). n = 1944 for S1P analysis and n = 1578 for S2P analysis

Identify factors that promote or limit human milk (HM) feeding and direct breastfeeding (BF) in infants with single-ventricle congenital heart disease at neonatal discharge in stage 1 palliation (S1P) and stage 2 palliation (S2P) (4–6 months of age)

Preoperative feeding practices, demographic characteristics, and social determinants of health, along with the feeding route at discharge from the first and second intervention periods, as well as the patient’s clinical course, were significantly associated with breastfeeding and human milk feeding (HM/BF) outcomes. Direct feeding before surgery was positively related to breastfeeding, while bottle or tube feeding was negatively related. Significant differences were observed based on race and type of health insurance. Additionally, the duration of hospital stays during the first intervention period showed a significant association with breastfeeding at discharge. The results varied across sites, suggesting the influence of site-specific factors on breastfeeding and human milk feeding outcomes

 He et al. (2006) Australia [27]

Supervised classification methods: decision tree and generalized linear model. Additionally, risk pattern mining is applied to identify cohorts at high risk of not breastfeeding

Prospective study including healthy mothers: n = 625 at hospital discharge; n = 544 at three months; and n = 372 at six months

To study the factors influencing the decision to breastfeed or not to a newborn

The study results show that applying data mining methods such as feature selection and supervised classification improves the accuracy in predicting the decision to feed the baby with breast milk. Classification accuracy on test data exceeds 75% when carefully selected features are used. Additionally, risk pattern mining identifies groups of mothers at high risk of not breastfeeding their babies, which could allow the implementation of targeted educational interventions to increase breastfeeding rates

Characterization of macronutrients in human milk

 Wong et al. (2021) Canada [28]

Machine learning models: Ordinary linear regression; Lasso regression by minimum angle; Random Forest regression; Gradient boosted decision tree regression

Breast milk samples from donor mothers to milk banks; n = 272

n = 61 mixed samples from multiple donors and n = 186 single donor samples. The fat and protein content of each individual donation and mixed sample was measured in duplicate using a mid-infrared human milk analyzer

Develop machine learning prediction models using donation-specific variables, maternal-infant characteristics, and milk expression variables to predict crude fat and protein content in donated milk samples

Machine learning models were able to predict crude protein levels in both individual and pooled samples with good accuracy (pooled milk with significantly lower error than baseline and clinically acceptable error), while fat prediction was more challenging due to its natural variability and measurement difficulty

 Jansen et al. (2010) [29]

Artificial neural network (ANN) approach combined with genetic algorithms (GA)

Human milk from three donors at the Human Milk Bank

Describe the use of an ANN-GA approach to characterize cholesteryl esters in human milk. Optimize high-performance liquid chromatography (HPLC) separation for cholesteryl ester characterization

Produced improved separation of nonpolar lipids compared to the pre-optimization method with ANNGA. The AI used allowed optimization of the analysis method using rpHPLC for efficient separation of cholesterol esters in complex biological samples like human milk without the need for prior purification steps. This will accelerate studies on the biological role of individual cholesterol esters in secretions

 Ruan et al. (2022) China [30]

Machine learning techniques with linear regression algorithms. | 546 milk samples from 244 Chinese mothers (from day 1 to day 1086 postpartum)

546 milk samples from 244 Chinese mothers (from day one to day 1086 postpartum)

Generate mathematical models to adjust macronutrient and energy values in human milk

Machine learning was used to adjust the results obtained by ultrasound and make them comparable to the values obtained by mid-infrared spectroscopy (MIR). Statistically significant differences were found between sample groups measured by MIR and ultrasound methods. Machine learning was used to generate mathematical models for three macronutrients (protein, fat, lactose) and human milk energy to adjust the results obtained by ultrasound and make them as close as possible to MIR values. After applying the adjustments using machine learning, the values generated by the two methods (MIR and ultrasound) were comparable and showed high consistency

Education and resolution of breastfeeding questions and issues

 Corrêa et al. (2023) Brazil [31]

Deep learning-based natural language processing (DL-based NLP Pipelines) to identify user intent in interactions with the chatbot and automatically classify text messages into different user intents. Chatbots were developed using a co-design approach. These chatbots were designed to function as virtual breastfeeding consultants using text messages and illustrative images through platforms like Telegram and WhatsApp

n = 18 health professionals specializing in breastfeeding working in a university hospital

Present the development process of Lhia (acronym in Portuguese for “Human Milk and Artificial Intelligence”), a chatbot focused on breastfeeding education and recruiting human milk donors

Focuses on evaluating and improving the chatbot’s performance throughout the co-design process, including intent identification, user engagement, number of interactions, model accuracy, and conversational flow quality. Training and validating Lhia based on feedback from health professionals achieved an accuracy of P3 with 0.82, P5 with 0.86, P5 with 0.81, and P4 with 0.93. This would increase breastfeeding rates, decrease early weaning, and increase milk donation. The accuracy of different NLP pipelines was evaluated in each co-design round. It was observed that the P4 pipeline, which uses BERTimbau word embedding, had the highest accuracy (0.93) and was selected for implementation in the production version of the chatbot. The chatbot’s conversational flow was improved based on suggestions and interactions from health professionals. An increase in the number of chatbot responses was recorded in each round, reflecting the expansion of conversational flow and improvement in the quality of responses provided

 Achtaich et al. (2023) [32]

1. Convolutional Neural Networks (CNN): to classify images of breasts related to breastfeeding and detect problems such as mastitis, sore nipples. Natural Language Processing (NLP): to understand and generate natural language responses. This NLP engine was used to process user text messages and generate relevant responses. 3. Artificial Intelligence Markup Language: A knowledge base was built using AIML to select appropriate responses to user requests. 4. Twilio API: The Twilio API was configured to enable bidirectional communication between the chatbot and users via the WhatsApp messaging platform. 5. Machine Learning: algorithms such as Naive Bayes (NB) and Support Vector Machine (SVM) to classify data and train models in the process of building the NLP engine and AIML knowledge base. | Availability of data collected from public online sources and not from the participation of specific individuals in the study itself

Availability of data collected from public online sources and not from the participation of specific individuals in the study itself

Develop the ALMA chatbot to support breastfeeding mothers by answering their questions related to breastfeeding and addressing their concerns about breastfeeding

Although the first version of ALMA was considered acceptable, areas for improvement were identified, particularly in expanding the knowledge base and implementing features to enhance empathy and voice support

 Oyebode et al. (2021) [33]

Various AI techniques were used to perform sentiment analysis on tweets related to breastfeeding: VADER (Valence Aware Dictionary and Sentiment Reasoner); TextBlob; Pattern; VADER-EXT (an extended version of VADER). Additionally, several machine learning algorithms were employed for text classification: Support Vector Machine (SVM); Multinomial Naïve Bayes (MNB); Stochastic Gradient Descent (SGD); Logistic Regression (LR); Random Forest (RF)

Sentiment analysis of tweets related to breastfeeding

The objective of this study is to determine a range of factors positively and negatively affecting breastfeeding behaviors

VADER-EXT was the best classifier, achieving an accuracy of 89.7% for negative polarity and a recall of 90.3% for positive polarity. SVM was the best classifier, with an accuracy of 74.02% and a recall of 73.95%. SGD also performed well

 Yadav et al. (2019) India [34]

Chatbot for breastfeeding education

Breastfeeding women and community health workers (ASHAs) in the East Delhi region, India

Understand the opportunities for chatbots in breastfeeding education for women in India

The study emphasizes the importance of addressing the barriers and challenges related to breastfeeding in contexts like India, where sociocultural and structural factors can significantly influence breastfeeding practices. Additionally, it highlights the potential of technologies such as chatbots to provide information and educational support to mothers and health professionals in this field

Detection of drug concentration and transfer in human milk

 Agatonovic-Kustrin et al. (2001) Australia [35]

- Genetic algorithm: Used to select a subset of molecular descriptors that best describe drug transfer to breast milk

- Artificial Neural Networks (ANN): Employed to correlate the selected descriptors with the M/P ratio and develop a QSAR (quantitative structure–activity relationship) model

123 structurally different compounds

Simplify and update the previously developed prediction model (neural network) for the milk-to-plasma (M/P) concentration ratio, based solely on the molecular structure of the drug

Identification of the most key factors influencing drug transfer to breast milk, including molecular size, shape, and electronic properties. The developed model does not require experimentally derived parameters and could provide a useful prediction of the M/P ratio for new drugs based solely on molecular structure, which could be valuable for drug information services. It should only be used as an aid in risk assessment along with the infant’s response

 Mashima et al

(2023)

Japan [36]

- Artificial Neural Networks (ANN)

- Support Vector Machine (SVM)

- Genetic Algorithm (GA)

M/P ratios were collected for 403 compounds, and M/PAUC was obtained for 173

Collect data on the relationship between drug concentration in human milk (M) and plasma (P) (M/P ratio) and build a binomial classification model based on the area under the curve (AUC) of the M/P ratio to detect drugs involved in active transport in mature human milk

The ANN model showed a sensitivity of 0.969 for the training set and 0.833 for the test set, with a specificity of 0.940 for the training set and 1.000 for the test set. The SVM model showed a sensitivity of 0.971 for the training set and 0.667 for the test set, with a specificity of 1.000 for both sets. The most influential descriptors were identified for each model, providing valuable information on the factors affecting drug transfer to human milk

 Zhao et al. (2005) [37]

- Ensemble learning technique: Boosting—Supervised learning methods: SVM (Support Vector Machines), LDA (Linear Discriminant Analysis)

126 common drug compounds

Develop dependable computational models to predict/classify drugs from milk to plasma (M/P). Develop SVM (Support Vector Machine) models to distinguish the potential risk of drugs for infants

The selected descriptors include characteristics such as polarity, molecular size, charge, geometry, binding energy, among others. The SVM model appears to be more effective in this specific context and might be preferable for predicting the risk of drug transfer to human milk compared to the LDA model. However, it is important to consider the specific characteristics of each model and the problem’s needs when selecting the most appropriate method

 Ye et al. (2022) China [38]

- Supervised learning: linear regression

- Image preprocessing and backpropagation artificial neural networks

Breast milk samples from two lactating mothers

Detection of amoxicillin in human milk

The results demonstrate that the proposed system is effective for the quantitative detection of amoxicillin in breast milk samples, with good selectivity and detection capability over a range of concentrations relevant for clinical applications

Detection of environmental contaminants in human milk

 Kowalski et al. (2013) Brazil [39]

Kohonen neural network

Colostrum, transitional, and mature milk from 193 Brazilian lactating women

Detect, identify, and quantify the presence of twelve PCBs in the breast milk of 193 lactating mothers from ten cities and towns across Brazil

Higher contamination was found in the milk of mothers living in large, industrialized cities, near polluted rivers or seas, in mature milk samples, and in mothers breastfeeding for the first time

 Nadal et al. (2004) Spain [40]

Artificial Neural Network (ANN); Self-Organizing Map (SOM) Kohonen

n = 54 human milk samples from different countries

The purpose of the study was to find a relationship between the profiles of different forms or chemical variants of persistent organic compounds of PCDD/PCDF in human milk and the dietary habits of the different countries where the samples were collected

The profiles of PCDD/F in human milk show high variability depending on the specific country or region. Countries with high fish consumption exhibit higher concentrations of PCDD/F in the milk. The SOM method is particularly relevant for providing an easy exploratory tool for cluster visualization

 Jovanovic et al. (2019) Serbia [41]

Guided regularized random forest with implementation of AutoWeka Metalearner

Milk samples from seventy-nine healthy mothers (23 primigravida’s, forty-one multigravidas, and fifteen multiparas)

Demonstrate a method to understand the relationship between organochlorine pesticides (OCP) and polychlorinated biphenyls (PCB) in breast milk and their association with the age and parity of the mother

This study revealed variations in POP levels in the milk of primiparas and multiparas. The ML methods provided relative prediction errors below 30% and correlation coefficients above 0.90, suggesting a possible nonlinear relationship between contaminants and the complexity of their pathways in breast milk