The human gut microbiome - a complex microbial community, its gene collection and metabolic potential - adapts dynamically across a lifetime, yet a critical period is the initial assembly in early life^[1-3]. It is initiated from birth onwards, with differential intrinsic and external exposures, including birth mode, feeding patterns and lifestyle factors shaping the colonisation trajectory^[4-8]. Mounting evidence suggests that these early microbial communities not only correlate with health conditions in infancy but may also distinguish between divergent health trajectories - current and future^[9-14]. While the former findings rely on statistical analysis, which remains indispensable for detecting relationships between microbial communities and host phenotypes, the latter are obtained by employing machine learning (ML), which extends analytical capacity from association toward prediction. ML is a branch of artificial intelligence capable of parsing complex data such as microbial metagenomes for meaningful patterns^[15]. Indeed, ML algorithms have been found successful in detecting microbiome-host interactions, predicting host phenotypes as well as potential disease risks based on microbial communities^[16-18]. It thereby serves as a tool for translating microbiome data into feasible clinical interventions used for potential microbial biomarker identification.

Several obstacles remain in employing operational microbiome-based predictive models in clinical practice. High inter-individual variability^[19], driven by genetic, environmental, and lifestyle factors^[20,21], has so far prevented the establishment of a benchmark ‘healthy’ microbiome^[22]. Without such reference points, defining optimal community structures or standardising risk prediction remains difficult. Moreover, while ML can identify prediction patterns linking microbial profiles to health outcomes, prediction does not equate to causality^[23], and the opaque nature of many algorithms raises concerns about interpretability^[24].

To date, most existing reviews in the field have focused on either microbiome-health associations in early life^[25-29], or on machine learning applications in adult populations^[16,23,24]. Given the central role of early development in shaping long-term health trajectories, and the limited effectiveness of many diagnostic tools in later childhood^[30,31], microbiome-based prediction in infancy holds considerable potential for earlier risk stratification, personalised intervention, and disease prevention. However, the early-life microbiome presents distinct analytical challenges, including rapid temporal dynamics, strong environmental sensitivity, and high inter-individual variability. These considerations in mind, a structured overview of how machine learning models perform in early-life cohorts, and how both methodological and biological factors influence predictive outcomes, is currently lacking.

This review addresses this gap by providing a focused evaluation of machine learning applications in the infant gut microbiome, with an emphasis on predictive modelling from birth to two years of age. In contrast to previous reviews, we critically assess how predictive models are constructed, which outcomes they target, and how performance varies across study designs and analytical frameworks. We further integrate key methodological aspects of microbiome research, including sequencing strategies and pre-processing pipelines, to evaluate their impact on model performance and cross-study comparability. By linking early-life microbial ecology with analytical and computational considerations, this review provides a structured framework for interpreting existing predictive evidence and clarifies the current limitations and requirements for developing robust, generalisable, and clinically meaningful microbiome-based prediction in infancy.

The initial stages of data acquisition and handling are essential for subsequent data quality and analysis. Microbial community profiling typically progresses from sample collection (e.g. saliva, stool or skin swabs) to DNA extraction, sequencing, preprocessing and data analysis. DNA extraction involves breaking open microbial cells using chemical, enzymatic, or mechanical methods to release DNA, followed by its purification from contaminants^[32]. Importantly, microbiome data are highly sensitive to variation in these pre-analytical steps. For example, DNA yield varies across stabilisation methods, time to freezing, freeze-thaw cycles, and storage time^[33-37]. In addition, DNA recovery can vary substantially based on extraction protocols, particularly due to lysis strategies and efficiency, which disproportionately affects Gram-positive bacteria with more resilient cell walls^[38-41].

Such differences in sample handling can introduce significant bias not only to DNA yield but also to inferred microbial profiles, which is especially critical in infant samples, considering their limited microbial community size^[37,38,42]. Sequencing requires the preparation of DNA libraries, by either amplifying specific marker regions or sequencing all genetic material^[43]. The two main approaches are whole-metagenomic shotgun sequencing (WGS) and 16S rRNA (16S) gene sequencing. The 16S gene, a genetic marker present in almost all bacteria, contains both conserved and variable regions, allowing taxonomic identification of bacteria and archaea typically up to the genus level, but without functional insight^[43-45]. For WGS, the genetic material found in samples is sliced into smaller parts (i.e. as if being shot with a gun), and all DNA is sequenced, thereby allowing sequencing of all microbes, including bacteria, archaea, fungi and viruses^[43,44]. While WGS enables higher taxonomic resolution and functional analysis, it is more computationally demanding, prone to host contamination and relies on less complete reference databases^{[20,43,44,46,47]}. Although decreasing cost makes WGS more accessible^[44], both approaches present challenges in sequence alignment and downstream interpretations.

The consequent processing and analysis of such complex and large data is challenging, both in terms of aligning the genetic sequences with reference databases and investigating associations between genetic material and disease phenotypes. While bioinformatic pipelines dealing with the former are extensively reviewed elsewhere^[20,46], we will focus on summarising challenges around the downstream analysis of microbiome data and attempts to combat them. Microbial communities underly principles of biological ecology, manifesting in complex relationships between microbes, and demanding consideration of such relationships in analysis^[22]. Moreover, interpersonal variability of compositional microbial profiles based on subjective microbial exposure makes for dataframes with large amounts of columns and zeros, i.e. sparse yet skewed feature distributions, which increases with larger sample sizes^[20,43,48]. Equally, large amounts of confounding variables ranging from technical considerations such as read counts or sequencing batches, study design factors including repeated measures, as well as covariates such as lifestyle and diet, further complicate microbiome data analysis^[48].

As such, a key feature of microbiome data analysis is to narrow down features, which also improves the interpretability of final outputs^[23]. These challenges are currently being tackled by uniquely creative solutions, which can broadly be divided into two categories, as shown in Figure 1: statistical and ML approaches. Statistical methodology aims to use exploratory approaches to answer specific hypotheses, most of which focus on microbiome differences at different developmental stages and between health and disease conditions. As such, statistical methods assess differences between individuals or groups in microbial community diversity or abundances of specific taxa, including compositional features or functional pathways. For the latter, referred to as differential abundance analysis, tests such as the Wilcoxon rank sum test or Fisher’s test, as well as a variety of software tools including ANCOM-BC2^[49], DESeq2^[50], or MaAsLin2^[51] are used, employing varying approaches such as mixed linear or binomial models. Lastly, taxa abundances can be correlated with any given clinical feature, providing a measure of the relationship between them. Taken together, statistical approaches provide interpretable differences between samples, quantifying group-level differences and indicating a degree of relatedness.

ML approaches, in turn, are able to handle greater amounts of more complex data, parsing it for underlying patterns and allowing statements regarding its predictive qualities. They are considered a type of artificial intelligence due to their intrinsic capacity to conduct trial and error processes, adjusting to outcomes automatically^[15]. Within ML, models can be categorised as supervised or unsupervised based on whether outcomes or predictors are set a priori^[15]. Unsupervised ML methods are predominantly used for dimensionality (i.e. feature) reduction, as well as clustering approaches such as k-means or hierarchical clustering^[16,23,24]. The goal of both is to identify how samples or features naturally group together, revealing underlying patterns or structures in the microbiome without prior knowledge or outcome labels.

Supervised ML methods in turn leverage a predefined set of factors to predict a specific continuous (regressor models) or categorical (classification models) outcome, making them a valuable tool for identifying clinical biomarkers or diagnosis^[15,23,48]. The respective workflow consists of several steps, starting with feature selection. Besides selecting the outcome variable of interest, a variety of dependent variables are tested for their predictive qualities, a process referred to as the training phase^[23]. This process is followed by a testing phase, where the selected features are presented with unseen data, resulting in a variety of model performance metrics^[23,43]. Subsequently, the model can be improved by an iterative flow through the previous steps, altering the selected features or model parameters to optimise predictive quality^[23].

Taken together, both statistical and ML approaches are essential and complementary in microbiome data analysis. While statistical methods are crucial in data preparation and characterisation, ML methodology allows for the recognition of patterns, clusters or predictive qualities of high-dimensional and non-linear microbiome data. As such, and especially in the light of the infant microbiome, ML presents an exciting avenue to extend analytical insight from association toward prediction of present and future health states. While prediction does not imply causation, integrating ML with longitudinal data and experimental validation strategies can help to move further towards causal relationships between microbial profiles and diseases.

The microbial ecosystem in the infant gastrointestinal tract begins to establish itself through a process referred to as microbial succession, whereby different species colonise the body in a relatively ordered yet individualised sequence, gradually building a stable and complex microbial community^{[1,13,19,52,53]}. Universally, infants experience initial colonisation at birth^[54], and the transmission process is thought to happen through both genetic predisposition and direct microbial exposure^[55]. As such, microbial colonisation is directed by perinatal factors such as individual gestational age and mode of delivery, reflected in vaginal births seeding microbes from the maternal vaginal and gastrointestinal tract, while infants born via cesarean section are primarily exposed to skin- and hospital-associated microbes^[1,4,56]. Accordingly, community structures of the two groups differ robustly in the first few months of infancy, with vaginally-born infants displaying higher numbers of species belonging to Lactobacillus, Bacteroides and Bifidobacterium, while infants born via cesarean section harbour higher amounts of Staphylococcus, Veillonella and Streptococcus species^[1,19,57].

Similarly, various factors, including antibiotic exposure, environmental microbial contact, and maternal health, as well as early-life feeding patterns, play key roles in the individuality of succession^[1,19,58]. While in general, microbial profiles of exclusively milk-fed infants are dominated by aerobic and facultative anaerobic taxa of Bifidobacteriaceae and Enterobacteriaceae^[8,52,59,60], individual differences based on varying amounts of either breastmilk or formula persist^{[1,6,13,57,61-64]}. Such dissimilarities are thought to trace back to breastmilk containing immunoregulatory compounds^[65,66] and promoting more limited amounts of species metabolising high amounts of prebiotic human milk oligosaccharides (HMOs) such as Bifidobacterium longum subsp. infantis^[67], while formula fosters a more diverse but less specialised microbiota^{[8,19,62,68-71]}. The introduction of solid foods, respectively, the stepwise cessation of breastmilk from around six months of age, encourages a major microbial transition^{[1,8,13,19,72]}. As a result, the presence of more anaerobic taxa capable of metabolising complex carbohydrates and proteins, such as Bacillota (formerly Firmicutes), Bacteroidota (formerly Bacteroidetes), and Clostridiales increases^[1,13,73]. Continuing throughout toddlerhood, higher abundances of species such as Faecalibacterium prausnitzii and Methanobrevibacter smithii, as well as increasing diversity and complexity of microbial community, result in mature, adult-like microbiota typically by age three years^{[1,19,55,74-77]}. As such, microbial succession follows predictable phases over the first two years of life, despite the various individual external and internal factors influencing the trajectory^[19,53,74].

Building on the understanding of microbial succession in early life, researchers have increasingly investigated how these dynamic microbial patterns relate to infant health outcomes. In most cases, such a role is established by statistically identifying factors associated with an outcome, followed by demonstrating differences in microbial profiles based on the same factors and the outcome, drawing an interactive effect triangle. As such, the early-life microbiome is thought to play a major role in healthy development and associated health and disease conditions, including, among others immuno-mediated diseases such as type 1 diabetes, inflammatory bowel disease, asthma, atopic dermatitis, and food allergies, as well as neurodevelopment, metabolic conditions, including obesity and gastrointestinal conditions such as necrotising enterocolitis (NEC)^{[27-29,78,79]}. Revealing these associations between the early-life microbiome and various paediatric disease states has fuelled interest in discovering microbial signatures that could serve as predictive biomarkers.

Early-life ML studies harness the predictive potential of the infant gut microbiome and can be classified into two broad categories, as summarised in Tables 1 and 2: predicting host phenotypes based on (groups of) individual microbial taxa or based on microbial maturation indices.

However, several recurrent limitations across the above-mentioned studies emerge, tempering the strength of their conclusions. Early-life microbiome data introduce a distinct set of analytical obstacles, primarily driven by their variability^[1]. Infant microbial communities are highly influenced by peri- and postnatal exposures, including birth mode, gestational age, antibiotic exposure, and rapidly evolving feeding practices^[13,112]. Driving substantial inter-individual variability, these factors make it difficult for models to learn stable patterns^[23,24]. Few of the studies adjust for a full range of these covariates^[74], making it difficult to disentangle microbiome-specific predictive signals from the influence of these wider determinants.

Another concern in prediction studies is overfitting due to the imbalance between high-dimensional microbial features and relatively small sample sizes^[23,48,113]. Indeed, several key findings, including predictions of urinary tract infections^[114], atopic dermatitis^[91], Type 1 diabetes^[92], behavioural outcomes^[95], and NEC^[84], are derived from cohorts with fewer than 150 samples, raising questions about the representativeness of their conclusions. While sample sizes are increasing, and several recent datasets now include thousands of samples, continued expansion of well-powered, multi-cohort studies will be essential to improve generalisability and reduce model variance.

Algorithm choice is another critical factor in infant microbiome modelling, as several of the reviewed studies report differing predictive performances and findings depending on the model applied^{[10,77,85,89]}. Infant microbiome data is compositional by nature, which violates the independence assumptions underlying many standard statistical models and can introduce spurious associations if not properly handled^[115]. This challenge is further compounded by the longitudinal nature of early-life sampling, where repeated measurements vary substantially across timepoints and between case-control groups. As a result, careful preprocessing, covariate adjustment, and model selection are essential to reduce noise and improve generalisability. As such, among classical ML methods, Random Forest has consistently demonstrated strong suitability for microbiome data^[16,115-118]. Random Forests are particularly advantageous due to their robustness to high-dimensional, sparse, and noisy data, minimal preprocessing requirements, and ability to provide intuitive measures of feature importance^[117,118]. Gradient boosting methods, including XGBoost and LightGBM, offer strong predictive performance by sequentially optimising models, and can capture complex non-linear relationships effectively^[77,118]. However, they are more sensitive to hyperparameter tuning, computationally intensive, and may be prone to overfitting in smaller datasets^[115]. As a result, their performance is often dataset-dependent rather than universally superior. More complex approaches, including deep learning models such as recurrent and convolutional neural networks, may be better suited for capturing temporal dynamics in infant microbiome data^[10,24]. However, these methods require substantially larger sample sizes and are computationally intensive, limiting their applicability in most current early-life studies. In practice, when data are limited, as is often the case in infant microbiome research, simpler ensemble methods tend to outperform deep learning approaches^[24]. However, performance differences between ML methods are often dataset-dependent, and no single algorithm consistently dominates across all microbiome applications.

Notably, across all model classes, interpretability remains a key limitation. While Random Forests decision trees provide some degree of transparency, more complex models, including gradient boosting and neural networks, are often considered ‘black boxes’^[115,119]. This poses a challenge for clinical translation, as identifying which taxa drive predictions is essential for biomarker discovery and potential pharmaceutical use. Post hoc interpretability tools, including SHapley Additive exPlanations (SHAP) values or permutation importance^[120,121], can partially address this issue, but they remain underutilised. Taken together, these considerations suggest that infant microbiome data should be handled with methods that account for compositionality, incorporate feature selection, and balance predictive performance with interpretability. Given current data constraints, ensemble tree-based methods represent robust and pragmatic choices, while more complex models may become increasingly relevant as larger, standardised datasets become available.

An additional critical and often overlooked limitation is inadequate validation. Across the studies reviewed, validation is almost exclusively restricted to internal or, in rare cases, external cohort-based evaluation. Internally validated models tend to produce overly optimistic performance estimates, with average intra-cohort performance reaching an AUC of 0.8 or higher, which typically declines substantially in external validation across independent cohorts^[16]. This reduction in performance likely reflects both biological variability (e.g. due to differences in geography^[52]) and technical heterogeneity, including differences in sample collection, sequencing platforms, and bioinformatic pipelines^[115]. Notably, none of the ML-based studies reviewed here incorporated experimental validation, such as in vivo or in vitro testing of model-derived microbial signatures. Only one study extended beyond prediction by performing downstream mechanistic investigations in necrotising enterocolitis^[88]. Casaburi and colleagues experimentally tested metabolite-driven hypotheses derived from metagenomic analyses in vitro and in vivo, demonstrating dose-dependent epithelial toxicity of candidate metabolites such as formate in rodent and human cell models. While this provides important mechanistic support for microbiome-associated disease signatures, it does not constitute validation of predictive ML models themselves, but rather offers biological plausibility for inferred microbial-metabolic associations^[88].

In summary, current early-life ML studies provide an important foundation for future translational research. The absence of independent and experimental follow-up limits causal inference, making it difficult to determine whether previously identified microbial signatures are drivers of disease or merely correlates of underlying pathology. As larger multi-cohort datasets, longitudinal study designs, and harmonised analytical frameworks continue to emerge, microbiome-derived signals are likely to become increasingly robust, reproducible, and clinically meaningful.

As such, rather than serving as immediate diagnostic tools, current predictive models should therefore be viewed as early-stage frameworks for identifying candidate risk patterns and guiding hypothesis generation. In the long term, integrating computational predictions with experimental validation and mechanistic studies will be essential for refining these signals into biologically grounded biomarkers. With continued methodological standardisation and expanding data availability, the early-life microbiome holds strong potential to support earlier risk stratification, improved disease prediction, and ultimately more targeted preventative strategies in paediatric healthcare.