A systematic review of machine learning on clinical MALDI-TOF MS

Abstract

Bacterial identification, antimicrobial resistance prediction, and strain typification are critical tasks in clinical microbiology, essential for guiding patient treatment and controlling the spread of infectious diseases. While machine learning (ML) has shown immense promise in enhancing Matrix-Assisted Laser Desorption/Ionization-Time of Flight Mass Spectrometry (MALDI-TOF MS) applications for these tasks, there is currently no comprehensive review that fully addresses this from a technical ML perspective. To address this gap, we systematically reviewed 115 studies published between 2004 and 2025, focusing on key ML aspects such as data size and balance, preprocessing pipelines, model selection and evaluation, open-source data, and code availability. Our analysis highlights the predominant use of classical ML models like Random Forest and Support Vector Machines, alongside emerging interest in deep learning approaches for handling complex, high-dimensional data. Despite significant progress, challenges such as inconsistent preprocessing workflows, reliance on black-box models, limited external validation, and insufficient open-source resources persist, hindering transparency, reproducibility, and broader adoption. This review offers actionable insights to enhance ML-driven bacterial diagnostics, advocating for standardized methodologies, greater transparency, and improved data accessibility. In addition, we provide guidelines on how to approach ML for MALDI-TOF MS analysis, helping researchers navigate key decisions in model development and evaluation.

Introduction

Identifying and characterizing microorganisms is essential in clinical microbiology to diagnose infections, guide effective antibiotic therapy, and support infection control [1]. Accurate and timely identification, along with determination of antimicrobial resistance (AMR) profiles through antibiotic susceptibility testing (AST), enables clinicians to select appropriate treatments and reduce hospital-acquired infections by up to 30 [2, 3]. However, standard AST methods—such as disk diffusion, broth microdilution, and automated systems—are resource-intensive and slow, often taking 24 to 72 h [4, 5], which can delay treatment decisions and hinder infection control. Beyond AMR, bacterial strain typification (TYP) is essential for disease management and epidemiological surveillance. TYP is based on molecular techniques such as polymerase chain reaction, pulsed-field gel electrophoresis, and multilocus sequence typing (MLST), although the field is increasingly shifting toward whole-genome sequencing (WGS) for its higher resolution. However, WGS remains costly, technically demanding, and often challenging to interpret, thereby limiting its routine use in diagnostic laboratories [6].

The introduction of Matrix-Assisted Laser Desorption/ Ionization-Time of Flight Mass Spectrometry (MALDI-TOF MS) in the late 1980s [7] revolutionized microbiology by enabling rapid, reliable, and cost-effective microorganism identification [8, 9]. By analyzing protein mass spectra, it provides species-level results in under 5 minutes, with a low cost per sample (0.5€/sample) and high-throughput capabilities, making it a standard in clinical laboratories [10].

The technique involves applying an organic matrix to the sample, enabling laser-induced ionization. The resulting ions are separated by their mass-to-charge (m/z) ratio in a time-of-flight analyzer, generating a high-dimensional spectrum that serves as a molecular fingerprint for each microorganism [11–13].

Recent advances in machine learning (ML) have shown great potential to automate microbial identification (ID) and AMR prediction from MALDI-TOF MS data [14–23]. Algorithms such as Support Vector Machines, Random Forests, and Neural Networks can extract complex patterns from spectra, reducing workload and enabling faster, more accurate diagnostics [24–26]. These tools also offer promising support for strain characterization and outbreak management [27–31].

However, challenges remain, particularly the need for large, high-quality datasets, standardized preprocessing, and solutions to class imbalance—especially in AMR and strain typing (TYP). Emerging approaches such as generative artificial intelligence (AI) and advanced deep learning (DL) models can help address these issues through data augmentation [32, 33], enhanced interpretability [34], and broader clinical applications. Initiatives like DRIAMS and MARISMa databases [21, 35, 36] highlight the importance of open-access spectral repositories to accelerate progress in the field.

While several reviews have highlighted the promise of ML in microbial diagnostics, most lack a strong technical perspective on model development and validation. For instance [5], focused on bacterial characterization and AMR prediction up to 2020, emphasizing species and antibiotic coverage but offering only limited discussion of ML workflows. [37] analyzed AMR prediction using ML between 2016 and 2024, but excluded ID and TYP tasks and overlooked key aspects such as data preprocessing, imbalance handling, and hyperparameter tuning. Similarly [38], adopted a broader scope on AI in antimicrobial stewardship beyond MALDI-TOF MS. Although valuable for microbiologists, these reviews fall short in addressing reproducibility, transparency, and ML-specific design considerations.

To bridge this gap, we present a systematic review of ML applications to MALDI-TOF MS for bacterial ID, AMR prediction, and strain TYP, covering studies from 2004 to 2025. We analyze the complete ML pipeline—from data preparation to clinical validation—with the goal of offering practical guidance for researchers in both microbiology and data science. While referencing [5] for pre-2020 studies, we apply stricter inclusion criteria, focusing exclusively on human bacterial pathogens and ensuring methodological rigor. Specifically, we examine the following:

• Trends in task scope, data size, instrumentation, and bacterial/antibiotic coverage.

• Exploratory data analysis, preprocessing, feature selection, and data imbalance handling, highlighting the lack of homogenous preprocessing pipelines.

• Model selection, highlighting black-box reliance and its impact on transparency, interpretability, and reproducibility.

• Evaluation metrics, cross-validation (CV), and external validation, identifying gaps in standardization.

• Open-source challenges in dataset and software accessibility, proposing strategies for greater transparency and reliability.

• Limitations of ML in MALDI-TOF MS and future directions for clinical integration.

In doing so, we propose the best practices to promote reproducible, interpretable, and clinically applicable ML workflows. This review serves as a methodological roadmap for the next generation of AI-enhanced microbiology. Figure 1 summarizes the structure of the paper.

Figure 1.

Graphical overview of the systematic review process on ML and DL applications to MALDI-TOF MS data, including (i) systematic search in Scopus, PubMed, and Web of Science, (ii) study selection based on eligibility criteria, (iii) data extraction covering research tasks, acquisition devices, bacterial species, databases, antibiotics, preprocessing steps, ML models, evaluation metrics, software, and open-source availability, (iv) summary of key limitations, and (v) definition of guidelines for future research.

Methods

Search strategy

A comprehensive literature search was conducted using PubMed/Medline, Scopus, and Web of Science, covering studies published from 1 January 2020 to 31 December 2025. The search aimed to identify studies that applied ML techniques to the analysis of MALDI-TOF MS data. The search included related terms¹ applied to titles, abstracts, and keywords when available. After retrieving 2616 documents and incorporating 36 earlier studies from Weis et al. [5], the study selection process was further refined by applying exclusion criteria to ensure the relevance and quality of the studies. The following exclusion criteria were applied:

• Duplicate records across repositories (PubMed/Medline, Scopus, and Web of Science) were removed during the initial screening phase.

• Irrelevant type records, such as review articles, letters, conference proceedings, or non-full-text articles in English, were excluded.

• Off-topic studies unrelated to MALDI-TOF MS and AI-based methods, such as research focused on genetics, proteomics, cancer, or other unrelated fields, were excluded.

• Articles focusing on nonhuman bacteria, such as those related to foodborne pathogens, animal studies, viruses, or fungi, were removed.

• Off-task studies that did not directly address ID, AMR, or TYP were excluded. Examples include studies assessing risk factors, mortality, morbidity prediction, or case studies.

The PRISMA flow diagram (Fig. 2) summarizes the screening process and the final set of included studies. This process led to a final selection of 115 studies published between 2004 and 2025. The overlap with previous reviews such as Weis et al. [5], López-Cortés et al. [39], and Pennisi et al. [38] was limited due to our stricter, ML-specific inclusion criteria.

Figure 2.

PRISMA 2020 flow diagram illustrating the literature search and screening process, showing the number of records identified from PubMed, Scopus, and Web of Science, records included from [5] ( = 36), and records screened and excluded at each stage based on eligibility criteria (duplicates, off-topic studies, nonhuman bacteria, and studies without AI-based methods), resulting in 115 included studies.

Data extraction

To ensure a comprehensive review, we extracted the following from each study: (i) publication details; (ii) the key research task (ID, AMR, or TYP); (iii) bacteria analyzed; (iv) antibiotics under study for AMR tasks; and (v) data description, including sample size and spectra details. For computational methodologies, we analyzed (vi) dataset details and feature importances; (vii) preprocessing and (viii) feature selection strategies; (ix) applied algorithms; and (x) evaluation metrics. Under the section model proceedings, we detailed (xi) hyperparameter tuning methods and (xii) the evaluation strategies, including CV and Out-of-Distribution (OOD) testing. Finally, (xiii) we assessed open-source data and code availability to evaluate transparency and reproducibility.

This systematic approach ensured consistency in the literature analysis and facilitated robust comparative analysis across all included studies. Tables 1, 2, and 3 summarize the extracted data.

Table 1.

Overview of included studies (2004–2025) categorized by year of publication, summarizing 115 studies analyzing MALDI-TOF MS data with Machine Learning (ML) approaches, where a single quotation mark (’) in the Isolates column indicates spectra, bold algorithms denote primary or best-performing models, an asterisk (*) in Preprocessing indicates steps subjected to ablation studies, citation counts (Cites) are from Google Scholar as of 26 January 2026, and notation for bacterial species, antibiotics, preprocessing steps, and algorithms is detailed in Tables 2 and 3.

Year	Paper	Task	Bacteria	Antibiotic	Isolates	Preprocessing	Peak Selection	Algorithm	Cites
2004	Zhang [40]	ID	Bacillus, S.au	-	69	N	-	RBF-MLP	29
	Satten [41]	ID	E.co, S.py, B.an, S.pn	-	120	S	I	RF	86
2008	Hsieh [42]	ID	S.au, E.co, K.pn, S.en, P.ae, GBS	-	132	N	-	QC, SVM, GA	264
2012	Schaumann [43]	AMR	E.co, K.pn, P.ae	ESBL, MBL	76	-	-	SVM	71
2013	Khot [44]	ID	Shigella, E.co	-	138	BaNA	O	GA	186
2014	Nakano [45]	AMR	E.fa	VAN	132	-	GA	GA, MLP, QC	50
	Li [46]	AMR	E.co, K.pn	CTX	103	T	SNR	GA, MLP, QC	18
	Xiao [47]	TYP	M.pn	-	68	BaNA	AUC	GA	41
2015	Zhang [48]	TYP	S.au	-	290	BaSNTA	AUC, SNR	GA	57
	Nakano [49]	TYP	S.pn	-	574	-	-	GA, MLP, QC	52
	Chen [50]	ID	S.pn, S.mi, S.or	-	75	-	S	GA, SVM, MLP, QC	60
2016	Camoez [51]	TYP	S.au	-	95	BaA	-	GA, MLP, QC	98
	Mather [52]	AMR	S.au	VAN	80	MQ	SNR	SVM	88
2017	Ho [53]	TYP	B.fr	-	424	-	AUC	GA, SVM, MLP, QC	39
	Sogawa [54]	AMR	S.au	MET	160	-	-	SVM	55
2018	Wang [55]	TYP	S.au	-	125	BaST	SNR, O	DT, SVM, KNN	73
	Wang [56]	TYP	S.au	-	306	BaS	S, GA	QC, MLP, KNN	54
	Schaumann [57]	ID	Clostridium	-	123	-	-	SVM	16
	Rodrigues [58]	TYP	Klebsiella	-	95	BaSTA	SNR	SVM	156
	Asakura [59]	AMR	S.au	VAN	129	BaSNBi	RF	RF	30
	Wang [60]	AMR	S.au	VAN	125	BaST	RF, SNR	SVM, KNN, DT, RF	60
2019	Ling [61]	ID	Shigella, E.co	-	23	-	-	MLP	17
	Chung [62]	TYP	S.ha	-	254	BaS	RF, AUC	LR, SVM, DT, RF	29
	Tang [63]	AMR	S.au	MET	20	MQ*	RF, O	BI, RF	57
	Wang [64]	TYP	GBS	-	787	BaSTBi*	SNR, O	RF, SVM	39
2020	Desaire [14]	ID	20	-	571	N	O	AC	18
	Pena [27]	TYP	K.pn	-	1056	BaSTP	SNR, O	SVM	10
	Huang [19]	AMR	K.pn	IPM, MEM		Ba	S	SVM, RF, KNN, NB, LR	54
	Ling [65]	ID	14	-	98 000	BaNT	SNR	MLP	5
	Kann [66]	TYP	S.pn	-	69	MQ	O	RF, CART	17
	Wei [67]	ID	B.an, B.ce	-	178	BaNTA	GA, SNR	GA	11
	Weis [25]	AMR	E.co, K.pn, S.au	AMC, CRO, CIP, TZP, PEN	2676	MQN*	-	LR, PIKE-GP	39
2021	Wang [68]	AMR	S.au	MET	4858	NBi	S, SNR	DT, RF, KNN, SVM	45
	Gato [69]	AMR	K.pn	MEM, ETP	1458	BaSA	-	PLS-DA, RF	15
	Mortier [15]	ID	1,000	-	89 767	BaNBi*	-	RF, LR, SVM, KNN, MLP	75
	Wang [70]	AMR	E.fa	VAN	7997	BaSTBi	S	RF, SVM, KNN	31
	Fiamanya [71]	TYP	B.co	-	55	BaSNTP	SNR	GA, MLP, QC	16
	Tang [72]	AMR	S.au	MET	960	-	GA	GA, MLP, QC	8
	Chung [73]	AMR	S.au	OXA, CLI, ERY		BaT*	RF, SNR	DT, SVM, RF	9
	Calderaro [28]	TYP	C.di	-	950	BaSTA	GA, SNR	GA, QC, MLP	25
	Huang [74]	TYP	GBS	-	1410	BaS	GA	KNN, MLP, QC	8
	Liu [75]	AMR	S.au	MET		MQBi	O	RBF-SVM, RF	27
	Calderaro [20]	AMR	E.co	COL	1040	BaSTAP	S, SNR	GA, QC, MLP	9
2022	Li [16]	ID	Listeria	-		BaSN*	O	KNN, NB, RF, SVM, MLP, DAE-SVM	34
	Feucherolles [76]	AMR	C.je, C.co	CIP, ERY, TCY, GEN, KAN, STR, AMP	340	BaNTA	RF, SNR	RF, LR, NB	62
	Weis [21]	AMR	803	42	300 000	MQBi	-	LR, LGBM, MLP	233
	Kong [77]	AMR	S.au	MET	548	NBi*	O	GA, SVM, DT, RF, PR	10
	Wang [78]	AMR	E.fa	VAN	7997	BaST	SNR	MLP, RF, XGB	37
	Wang [79]	AMR	K.pn	CIP	15 782	BaSTBi	S, SNR, O	LR, SVM, RF, XGB	22
	Dematheis [80]	ID	B.me, B.ab, B.su	-	44	MQA	RF	RF, SVM, MLP, MNR	23
	van den Beld [81]	ID	Shigella, E.co	-	559	BaSA	O	SVM	20
	Jeon [82]	AMR	S.au	MET	553	BaSNABi	RF	RF	20
2022	Candela [83]	AMR	E.fa	VAN	178	BaSNA*	-	SVM, RF, PLS-DA	33
	Zhang [84]	AMR	S.au	OXA, CLI	31 815	BaSTBi	SNR, O	LR, XGB	18
	Wang [85]	AMR	K.pn	IPM, MEM, ETP	171	-	RF	RF, SVM, SVM-k	18
	Yu [86]	AMR	S.au	MET	20 359	BaSNBi*	-	LGBM, GB, LR, XGB, ET, RF, SVM, DT, KNN, NB	54
	R.-Villodres [87]	AMR	E.co	TZP	248	BaSNTA*	SNR	PLS-DA, SVM, RF	6
2023	Wang [88]	AMR	S.au	PEN, OXA, ERY, CLI, SXT, FUS	31 807	BaNTBi	SNR, O	XGB	8
	G.-López [89]	AMR	K.pn	15	402	MQ*	O	KSSHIBA-FA, RF, SVM, PIKE-GP	18
	Candela [17]	ID	E.cl	-	337	BaSNA	-	PLS-DA, SVM-L, SVM-R, RF	22
	Chung [90]	AMR	E.co	AMX, CAZ, CIP, CRO, CXM	37 918	BaST	RF, SNR	LR, SVM, RF, XGB	9
	Gato [91]	AMR	K.pn	MEM, ETP	715	BaSA*	-	SVM, PLS-DA, KNN, RF	21
	Yu [92]	AMR	K.pn	IPM, MEM, ETP, COL	3938	BaSNBi	-	LGBM	35
	Abdrabou [29]	TYP	C.di	-	240	BaSNA*	I	SVM, PLS-DA, KNN, RF	16
	Zintgraff [93]	TYP	S.pn	-	238	MQAPBi*	O	GLMNET, RF, NB, KNN, LDA, MLP, SVM, DT	3
	R.-Temporal [94]	ID	M.ab	-	325	BaSN	RF	PLS-DA, SVM, RF, KNN	35
	Zeng [22]	AMR	K.pn	IPM	174	BaSNBi	-	LASSO, LR, SVM, MLP	7
	Yu [95]	AMR	K.pn, S.au	IPM, MEM, ETP, MET	170	BaSNBi	-	LGBM	20
	Wang [96]	TYP	S.si	-	205	MQ	-	RF, SVM	4
	Fu [97]	AMR	P.ae	TOB, FEP, MEM		MQBi	RF	MLP	5
	Manfredi [30]	TYP	E.co	-	202	BaS	S	MLP, GA-KNN, QC	13
	Wang [98]	ID	M.ab	-	102	Bi	RF	RF, SVM, DT, LR, KNN	14
	Iskender [99]	AMR	K.pn	COL	260	BaN	GA	LDA, SVM	5
	Liu [100]	ID	B.lo, B.in	-	100	SN	O	RF, SVM, LR	7
	Candela [101]	TYP	P.ae	-	67	BaSNA	O	PLS-DA, SV, RF, NCA-KNN	7
	Zhang [102]	AMR	K.pn	IPM, MEM, DOR, ETP	2683	MQBi	O	MLP	28
	Chung [103]	AMR	A.ba, A.no, E.fa, GBS	CIP, VAN, CLI	22 101	MQBaS*	RF	LR, NB, RF, SVM	12
2024	Wang [104]	AMR	S.au	OXA, ERY, CLI, TCY, SXT, MET	5290 ’	BaSNBi	O	LGBM, MLP, RF, KNN, AdaB, LR	5
	de Waele [105]	AMR	25	74	55 773 ’	BaSNBi	-	MLP, LR, XGB	9
	Candela [106]	TYP	C.di	-	69	BaSNA	-	KNN, SVM, PLS-DA, RF, LGBM	3
	Lin [107]	AMR	K.pn	CPA	675	NBi	RF	LR, LDA, RF, GB, AdaB, XGB, LGBM	9
	Jian [108]	AMR	K.pn	COL, DOR, IPM	4307	NBi	O	LDA, RF, GB, AdaB, XGB, LGBM, SVM	9
	Huang [109]	TYP	GBS	-	167	A	GA	GA, MLP, QC, KNN	5
	Godmer [31]	TYP	C.di	-	201	MQ	SNR	MNR, MLP, RF, SVM	10
	Mao [18]	ID	E.co	-	38 400 ’	T	SNR	LSTM	9
	Lin [23]	AMR	S.ma	LVX, SXT	7266	NBi	RF	LR, LDA, AdaB, LGBM, XGB, RF, GB	8
	Xu [110]	AMR	K.pn	ETP, IPM, MEM	240	-	O	RF, SVM, LR, XGB	3
	L.-Cortés [39]	AMR	E.co, K.pn, S.au	CIP, CRO, FEP, TZP, TOB, MEM, FUS, OXA	11 600 ’	BaSBi	-	MLP	28
	Astudillo [111]	AMR	E.co, K.pn, S.au, P.ae	OXA, CLI, FUS, CIP, CRO, TZP, FEP, IPM, MEM	13 157 ’	Bi	-	MLP, SVM, RF, XGB	11
	Jian [112]	AMR	K.pn	LVX, CIP	11 996	NBi	O	LR, LDA, RF, GB, AdaB, XGB, LGBM	10
	Macaya [113]	AMR	E.co, K.pn, S.au	CRO, OXA	30 069	NBi	-	MLP	5
	Nguyen [114]	AMR	P.au	11	360	MQBi*	-	LR, RF, SVM, LGBM, MLP, T	23
	Ren [115]	AMR	S.ep	20	75 042 ’	MQBi	-	RF, LR, NB, SVM, LGBM, MLP	16
2025	Nishikawa [116]	ID	K.pn	-	39	MQBi	O	BI	1
	Candela [117]	TYP	C.di	-	69	BaSNA	AUC	RF, LGBM, PLS-DA, SVM, KNN	3
	Hmidi [118]	TYP	C.di	-	389	BaSN	-	SVM, RF, KNN, GLM, LDA, GB, MLP, GP	0
	Lin [119]	AMR	K.pn, P.ae	AMK, GEN, TZP, CAZ, CRO, TZP, FLO, FEP, IPM, DOR, CIP, LVX	22,396	NBi*	RF	LGBM, RF	2
	de Waele [120]	AMR	469	65		BaSN	I	MLP, RF, LR, KNN, T	7
	Godmer [121]	ID	E.co, M.ab, S.pn	-	414	BaSNT*	SNR, RF, O	MNR, RF, SVM, XGB, MLP	2
	Duque [122]	ID	Morganella	-	235	BaSNTBi	RF	MNR, RF, SVM, MLP	1
	Cai [123]	AMR	K.pn	IPM, ETP, MEM	1,532	-	O	RF, SVM, DT, NB, LR	0
	Xu [124]	AMR	K.pn	TZP, CAZ, CRO, CTT, ATM, IPM, AMK, LVX, SXT	484	SBi	-	RF, XGB, AdaB, LR, MLP, SVM	0
	Yong [125]	AMR	S.au	MET	1,464	MQBi	-	LR	1
	Nithimongkolchai [126]	AMR	B.ps	CAZ, MEM, IPM, AMC, SXT	300	MQBi	-	DT	1
	Lin [127]	AMR, TYP	S.au	MET	10,227	NBi	RF	RF, LGBM, XGB	1
	Lu [128]	AMR	K.pn	CIP, CXM, CRO	28,000	MQ	O	NB, DT, LR, SVM, RF, XGB	0
	Ye [129]	TYP	K.pn	-	205	BaSN	-	RF, SVM, AdaB, MLP	0
	Ye [130]	AMR, TYP	K.pn	IPM, ETP, MEM	267	-	-	GA, MLP, QC	1
	Ren [131]	TYP	Salmonella	-	692	BaSNT	SNR, O	AdaB, MLP, DT, LGBM, KNN, GB, RF, SVM, XGB, NB	1
	Wiesmann [132]	AMR	E.co, K.pn, S.au	AMP, CIP, CTX, SXT, TZP, PEN, OXA	15,005	MQBi	-	LR, MLP, SVM, RF, LGBM, XGB	1
	Wang [133]	TYP	C.di	-	385	MQBi	S	MLP	0
	B.-Sánchez [134]	TYP	C.di	-	379	BaSN	O	KNN, SVM, RF, LGBM, DBLR, FA-VAE	1
	Xu [135]	AMR	E.co, K.pn	IPM, ETP, MEM	1,084	SBi	-	DT, RF, GB, XGB, ET	4
	Thai [136]	AMR	M.ab	AMK, LNZ, CLR, FOX	50	MQBi	RF	DT	2
	L.-Cortés [137]	AMR	E.co, K.pn, S.au	OXA, CIP		MQBi	O	MLP	9

Table 2.

List of codes and abbreviations used for bacterial species and antibiotics.

Code	Description	Code	Description
Bacterial species
A.ba	Acinetobacter baumannii	E.co	Escherichia coli
A.no	Acinetobacter nosocomialis	GBS	Group B Streptococcus
B.an	Bacillus anthracis	K.pn	Klebsiella pneumoniae
B.ce	Bacillus cereus	M.ab	Mycobacterium abscessus
B.fr	Bacteroides fragilis	M.pn	Mycoplasma pneumoniae
B.in	Bifidobacterium infantis	P.ae	Pseudomonas aeruginosa
B.lo	Bifidobacterium longum	S.au	Staphylococcus aureus
B.ab	Brucella abortus	S.ep	Staphylococcus epidermidis
B.me	Brucella melitensis	S.en	Salmonella enterica
B.su	Brucella suis	S.ha	Staphylococcus haemolyticus
B.co	Burkholderia contaminans	S.ma	Stenotrophomonas maltophilia
B.ps	Burkholderia pseudomallei	S.mi	Streptococcus mitis
C.co	Campylobacter coli	S.or	Streptococcus oralis
C.je	Campylobacter jejuni	S.pn	Streptococcus pneumoniae
C.di	Clostridium difficile	S.py	Streptococcus pyogenes
E.cl	Enterobacter cloacae	S.si	Streptococcus suis
E.fa	Enterococcus faecium
Antibiotics
AMC	Amoxicillin/Clavulanic acid	FLO	Flomoxef
AMK	Amikacin	FOX	Cefoxitin
AMP	Ampicillin	FUS	Fusidic acid
AMX	Amoxicillin	GEN	Gentamicin
ATM	Aztreonam	IPM	Imipenem
CAZ	Ceftazidime	KAN	Kanamycin
CIP	Ciprofloxacin	LNZ	Linezolid
CLI	Clindamycin	LVX	Levofloxacin
CLR	Clarithromycin	MBL	Metallo -lactamases
COL	Colistin	MEM	Meropenem
CPA	Ceftazidime-avibactam	MET	Methicillin
CRO	Ceftriaxone	OXA	Oxacillin
CTX	Cefotaxime	PEN	Penicillin
CTT	Cefotetan	STR	Streptomycin
CXM	Cefuroxime	SXT	Trimethoprim/Sulfamethoxazole
DOR	Doripenem	TCY	Tetracycline
ERY	Erythromycin	TOB	Tobramycin
ESBL	Extended spectrum -lactamases	TZP	Piperacillin/Tazobactam
FEP	Cefepime	VAN	Vancomycin

Table 3.

List of codes and abbreviations used for preprocessing and feature selection steps, as well as machine learning (ML) models.

Code	Description	Code	Description
Preprocessing steps
MQ	MALDIquant [138]	Ba	Baseline subtraction
S	Smoothing	N	Normalization
T	SNR thresholding	A	Averaging
Bi	Binning
PS techniques
AUC	Area under curve	GA	Genetic algorithm
I	Intensity thresholding	RF	Random forest
S	Statistical tests	SNR	Signal-to-noise ratio
O	Other
ML algorithms
AC	Aristotle Classifier	LDA	Linear Discriminant Analysis
AdaB	AdaBoost	LGBM	Light GB Machine
BI	Binary Discriminant Analysis	LR	Logistic Regression
CART	Classification Regression Tree	LSTM	Long Short-Term Memory
DAE	Denoising Autoencoder	MLP	Multi-Layer Perceptron
DBLR	Dual Bayesian LR	MNR	Multinomial Logistic Regression
DT	Decision Tree	NB	Naïve Bayes
ET	Extra Trees	PLS-DA	Partial Least Squares Discriminant Analysis
FA	Factor Analysis	PR	Polynomial Regression
FA-VAE	FA Variational Autoencoder	QC	Quick Classifier
GB	Gradient Boosting	RBF	Radial Basis Function
GLM	Generalized Linear Model	RF	Random Forest
GP	Gaussian Process	SVM	Support Vector Machine
KNN	k-Nearest Neighbors	T	Transformer
LASSO	Least Absolute Shrinkage and Selection Operator	XGB	Extreme GB

Quality assessment

Based on the data extracted from each study, following section Data extraction, we assess its quality according to the following key aspects:

• ML methodology: clear description of the algorithms used, including model architectures, hyperparameter validation ranges, and training procedures.

• Model evaluation: adequate evaluation metrics and validation strategies to ensure robustness and generalization, such as k-fold inner CV or the use of external datasets for OOD evaluation.

• Data transparency: availability of datasets and/or code to enable reproducibility.

• Clinical relevance: the practical application of the findings, particularly in ID and TYP bacterial species and predicting AMR in clinical settings.

Each of these aspects will be explored in detail in the Results section, while broader implications and key challenges will be discussed in the Challenges section.

Results

This section presents the main findings from the reviewed studies (=115²).

Key research task

Firstly, in ML it is important to distinguish between multi-class, multi-label, and multi-task problems. Multi-class classification assigns one label from several possible classes (e.g. predicting a single bacterial species). Multi-label classification allows multiple binary labels per sample (e.g. resistance to several antibiotics). Multi-task learning combines multiple prediction objectives—such as jointly performing multi-class ID and multi-label AMR prediction—within a single model.

Only a relatively small proportion of the reviewed studies focus on bacterial species ID (=22). This limited representation is expected, as MALDI-TOF MS is already a gold-standard technique for ID in routine clinical microbiology, often achieving high accuracy without requiring ML-based methods. Consequently, ML is typically applied in ID tasks only when addressing particularly challenging scenarios, such as differentiating closely related species or genera. Examples include species within the Morganella genus [122] or discrimination between phylogenetically similar organisms like Escherichia coli and Shigella [30, 139]. These studies highlight that, while conventional MALDI-TOF MS workflows perform well in standard settings, ML can provide additional resolution in complex classification problems.

AMR remains the dominant research focus across the literature, accounting for =64 studies. Most works aim to predict resistance profiles to specific antibiotics, reflecting the strong clinical demand for rapid susceptibility assessment. Among studies addressing multiple AMR profiles for bacterial species is typically perform binary classification by training separate models for each resistance phenotype (=24), the majority do not adopt multi-label learning strategies [21, 23, 25, 39, 73, 76, 88–90, 92, 97, 103, 104, 108, 112–115, 124, 126, 128, 132, 136, 137]. Only =5 studies employing genuine multi-label classification frameworks [84, 105, 111, 119, 120]. Notably, only one study simultaneously addresses bacterial ID and AMR prediction within a unified modeling framework [105], underscoring the limited exploration of multi-task approaches.

Strain typing (TYP) is addressed in =31 studies. While historically less prevalent than AMR, a notable increase in TYP-focused publications has been observed in 2025, where such studies account for 36% of the papers published, compared with roughly 25% in earlier periods. This recent growth suggests a renewed interest in bacterial strain characterization, likely driven by the increasing importance of epidemiological surveillance [131], outbreak detection [117, 134], and the need to differentiate strains exhibiting distinct resistance profiles [127, 129].

Summary. Task distribution across the reviewed studies reflects the clinical maturity of MALDI-TOF MS for bacterial ID, where ML is primarily used for challenging discrimination scenarios rather than routine ID. AMR prediction remains the predominant application, consistent with its direct impact on patient management. In contrast, the recent increase in TYP-focused research indicates growing recognition of the importance of strain-level characterization. Figure 3 shows trends in researched tasks over the years. Despite advances, multi-label and multi-task learning strategies remain underexplored, representing a key opportunity for improving the efficiency and clinical applicability of ML-based MALDI-TOF MS workflows.

Figure 3.

Overview of tasks addressed (top) and the best-performing ML models used (bottom) per year across studies, where in studies (=19) reporting two models as equally best-performing (see bold in Algorithm in Table 1), each model contributes 0.5 to the counts.

Data description

Before applying any ML model, it is essential to understand the dataset, particularly distinguishing between bacterial isolates and mass spectra. Isolates represent cultured strains from clinical or environmental sources, while spectra are their corresponding protein fingerprints obtained through MALDI-TOF MS analysis [140, 141]. One isolate may yield multiple spectra due to technical or biological replicates. However, studies vary in reporting—some specify isolates [75], others only spectra [16], or both [29]—leading to inconsistencies. Since isolates are the most relevant unit for ML, the lack of standardized data reporting complicates study comparisons.

Dataset sizes in the reviewed studies vary significantly, ranging from just 44 isolates [27] to over 300 000 [21, 120]. The majority (=69) rely on small datasets with fewer than 500 isolates, while =18 use medium-sized datasets (500–10 000 isolates), and only =13 incorporate >10 000 samples. The median dataset size is 244 isolates and 1429 spectra.

Data imbalance is a frequent issue, with less than one-third of studies (=32) working with balanced datasets [18–20, 22, 41, 42, 45, 46, 51, 53, 54, 63, 65, 68, 70, 72–74, 78, 82, 83, 85–87, 91, 93, 98–100, 103, 116, 123], while =83 report imbalances, particularly in AMR and TYP tasks. Notably, =10 studies fail to disclose balance status [29, 30, 40, 43, 55, 56, 60, 61, 96, 105]. To address this, =24 studies apply resampling methods [17, 23, 25, 31, 64, 79, 80, 84, 89, 94, 97, 107, 108, 111–113, 118, 119, 121, 124, 132, 134, 135, 137], including SMOTE [142] (=16) [31, 79, 97, 111, 113, 121, 124, 135, 137]. Others (=14) use imbalance-aware evaluation metrics such as area under the precision-recall curve (AUPRC) or balanced accuracy [16, 21, 25, 31, 39, 76, 89, 97, 101, 102, 104, 111, 113, 115, 120, 124, 125, 128, 133, 134, 137], while only =4 adjust loss functions [25, 110, 114, 132]. It is concerning that =35 acknowledge working with imbalanced datasets without applying corrective techniques [14, 15, 27, 28, 44, 47–50, 52, 57–59, 62, 66, 67, 69, 71, 75, 77, 81, 88, 90, 92, 95, 106, 109, 117, 122, 126, 127, 129–131, 136]. Dataset imbalance is an inherent characteristic of computational microbiology datasets, reflecting the natural distribution of species, strains, and resistance phenotypes. However, transparent reporting of class distributions and explicit consideration of imbalance during model development remain crucial, or otherwise require clear methodological justification.

The reviewed studies employ various MALDI-TOF MS data acquisition devices. The main manufacturer is Bruker Daltonics [143] (=89), preferred due to its high-throughput capabilities and well-established software tools, such as flexAnalysis and ClinProTools, which facilitate data analysis and integration with ML algorithms. The VITEK MS system by BioMérieux [144] (=16) lacks the widespread adoption of research seen with Bruker systems, although both provide a full pipeline from measurement to species identification via database lookup. In addition, =10 studies make use of less common or custom MALDI-TOF instruments, such as Zybio Inc., Applied Biosystems, PerSeptive Biosystems, Absciex, ASTA MicroIDSys, or Voyager DE-STR. Three studies [104, 121, 123] use data acquired from more than one manufacturer.

Summary. Establishing systematic data handling and reporting practices—explicitly stating dataset size, biological and technical replicates, and their management—is crucial to prevent bias, misinterpretation, and irreproducibility, ensuring scientific rigor. The lack of medium-large datasets remains a major challenge, though large repositories are beginning to mitigate this issue. Smaller datasets, while easier to handle, often hinder model reliability and generalizability, especially without robust external validation. Dataset imbalance emerges as a pervasive and inherently expected characteristic of computational microbiology data; however, mitigation and evaluation strategies are applied inconsistently, with several studies acknowledging imbalance without corrective measures. Finally, MALDI-TOF MS data acquisition is strongly dominated by Bruker Daltonics systems; while new manufacturers have entered the market in recent years, their adoption remains regionally limited.

Bacteria

The reviewed studies primarily focus on a few bacterial species that are clinically significant due to their association with hospital-acquired infections and AMR. The most prevalent species studied is Klebsiella pneumoniae (=31), reflecting its critical role in carbapenem-resistant infections, which pose a major challenge in healthcare settings [145]. Following closely, Staphylococcus aureus (=30), which is extensively studied particularly for the detection of methicillin-resistant S. aureus (MRSA), a notorious pathogen responsible for numerous healthcare-associated infections [146, 147]. Escherichia coli (=20), known for its role in urinary tract infections and rising resistance to antibiotics such as fluoroquinolones [148] and carbapenems [149], is also a frequent subject of analysis.

The majority of studies (=88) focus on bacteria from a single genus, such as Streptococcus [66, 74, 93, 96]. In contrast, a smaller subset (=27) examines bacteria across multiple genera, of which only =8 address more than three genera. This tendency likely stems from the focus of most AMR-related studies on single-species bacterial analysis for susceptibility to one or multiple antibiotics (see Section Antibiotics under study).

Summary. Although the diversity of studied bacteria is broad, covering 41 specified species, research efforts remain heavily concentrated on a limited subset of bacterial species, with 60% of studies focusing on S. aureus, K. pneumoniae, and E. coli. While this distribution is partly justified by their clinical relevance, its concentration introduces a potential research bias, as methodological advances are repeatedly validated on a small number of organisms, which may not capture the full biological and technical variability inherent to MALDI-TOF MS data. Consequently, reported performance levels may overestimate model robustness and generalizability. Since the release of DRIAMS [21, 35], several studies have incorporated a substantially larger number of bacterial species within a single experimental framework. This shift enables evaluation under more heterogeneous and clinically realistic conditions, opening the door to the development of more generalizable models and more complex learning paradigms. Nevertheless, most investigations still rely on single-task and multi-class formulations, limiting the ability of ML systems to fully capture the multi-label nature of clinical microbiology problems.

Antibiotics under study

The reviewed studies examine a diverse range of antibiotics to predict AMR across various bacterial species, encompassing a total of 39 different antibiotics and three antibiotic groups—namely carbapenems, extended spectrum -lactamases, and metallo -lactamases. The scope of antibiotics analyzed per study varies significantly; while =34 studies focus on a single antibiotic or antibiotic-family (e.g. carbapenems), others (=30) investigate resistance across a broader spectrum. Notably, the largest number of antibiotics examined in a single study reaches 74 [105]. As shown in Table 1, the number of antibiotics analyzed per study as well as the number of studies about AMR has increased over time, underscoring the growing importance of addressing multiresistant bacteria.

Among the antibiotics studied, carbapenems (CP) are the most frequently analyzed (=25), including meropenem (=15), imipenem (=15), ertapenem (=10), and doripenem (=3). Their critical role in treating infections caused by multidrug-resistant bacteria makes them a primary focus of AMR research. However, the growing threat of CP-resistant bacteria poses a significant public health threat [150] and has led many studies to classify resistance broadly as CP-resistant versus susceptible (=14), and others differentiating between specific CP drugs (=11).

Beyond CP, other commonly studied antibiotics include methicillin (=12), primarily in the context of MRSA detection [54, 63, 68, 72, 75, 77, 82, 86, 95, 104, 125, 127], and vancomycin (=8), widely analyzed for its role in Enterococcus faecium and vancomycin-resistant Enterococci (VRE) [45, 70, 78, 83, 103]. Additionally, ciprofloxacin (CIP) (=12) is frequently studied in E. coli and K. pneumoniae for urinary tract and systemic infections [25, 39, 79, 90, 111, 112, 119, 128, 132, 137]. Oxacillin (OXA) (=9) is consistently examined for its relevance in detecting S. aureus resistance, while clindamycin (CLI) (=6) is another frequently assessed antibiotic. Lastly, colistin (COL) (=4), a last-resort antibiotic, is commonly investigated for resistance in K. pneumoniae and other Gram-negative bacteria [20, 92, 99, 108].

However, the inclusion of multiple antibiotics within a single study does not necessarily imply that AMR prediction is performed jointly for all drugs within a unified modeling framework. In most cases, resistance prediction is still addressed through independent binary classification models. Only =5 studies adopt multi-label classification strategies for AMR prediction. [84] predict S. aureus resistance to two antibiotics [111], perform species-level multi-label classification across nine antibiotics and four bacterial species [105], investigate three modeling strategies: a single model for all antibiotics and species, individual models per species, and species–antibiotic pair models. [119] analyze resistance patterns for multiple antibiotics across two bacterial species, and [120] employ a recommender system framework to predict resistance profiles across multiple species and antibiotics.

The World Health Organization Bacterial Priority Pathogens List (WHO BPPL) has served as a tool for guiding research and development (R&D) investments, as well as shaping surveillance and control strategies for AMR [151]. The 2024 BPPL update categorizes antibiotic-resistant bacteria into three priority levels: critical, high, and medium. Analyzing how the literature aligns with these categories, it can be observed that for the critical-priority group where it is included CP-resistant Acinetobacter baumannii and Enterobacterales, such as K. pneumoniae and E. coli, in cases of CP and third-generation cephalosporin resistance (e.g. CTX, CAZ, CPA, CRO), we find =0, =14 and =13, respectively. In the high-priority category, there are studies for CP-resistant Pseudomonas aeruginosa (=4), MRSA (=12), VRE (=5), and fluoroquinolone-resistant Salmonella and Shigella species (e.g. CIP, LVX) (=0), all of which pose increasing threats due to their resistance profiles and prevalence in healthcare settings. While for the medium-priority group, there is no study for macrolide-resistant Streptococcus pneumoniae and Group A Streptococci (e.g. ERY resistance) and penicillin-resistant GBS, which represents growing risks, particularly in low- and middle-income countries. These gaps highlight the need for further ML-driven research tailored to high-burden-resistant bacteria, aligning with global AMR mitigation efforts.

Summary. AMR prediction has emerged as the dominant application of ML in MALDI-TOF MS, progressively displacing other tasks such as bacterial identification. However, most studies formulate resistance prediction as independent binary classification problems, a simplification that does not fully capture the multi-drug and multi-mechanism nature of clinical resistance. While this approach facilitates model development, it may limit clinical applicability, where simultaneous resistance profiling is often required. Notably, several high-priority pathogens and resistance scenarios highlighted in the 2024 WHO BPPL, including CP-resistant A. baumannii, remain underexplored or absent. Furthermore, modeling strategies better aligned with clinical reality, such as multi-label approaches, remain surprisingly scarce. Expanding beyond binary formulations and improving coverage of clinically relevant resistance threats represent key opportunities for future research.

Exploratory data analysis and feature importance analysis

Exploratory data analysis (EDA) plays a crucial role in understanding the structure, distribution, and variability of the data and helps identify patterns, inconsistencies, and outliers. In fact, it can help us select appropriate data preprocessing and adequate ML algorithms. Visualization is an essential aspect of EDA, as it enables a more intuitive interpretation of complex data structures, making it easier to uncover meaningful relationships and trends.

A total of =73 studies incorporate some form of EDA, with dimensionality reduction being the most common approach. Principal component analysis (PCA) is the most frequently used technique (=28) [15, 17, 18, 20, 25, 27, 29, 30, 47, 52, 55, 67, 71, 81, 83, 84, 87, 91, 93, 94, 96, 100, 101, 106, 111, 117, 120, 131], followed by t-distributed stochastic neighbor embedding (t-SNE) (=7) [17, 89, 90, 98, 114, 128, 131]. While these techniques are often used for dimensionality reduction to two or three dimensions, their visual representations can also provide valuable insights for data interpretation. Clustering methods are used to group bacterial strains or resistance profiles based on similarity, providing valuable insights into potential subgroups within the dataset. Such techniques include unsupervised hierarchical clustering analysis (UHCA) (=17) [17, 27, 42, 57, 63, 64, 66, 70, 80, 81, 91, 93, 94, 96, 101, 116, 117], dendrograms (=10) [28, 43, 45, 48, 52, 71, 81, 93, 96, 117], and k-means clustering (=2) [80, 110], aiding in pattern recognition and data exploration.

When working with very high-dimensional data, feature importance analysis becomes essential for understanding how different variables contribute to predictions and overall model performance. Identifying the most influential features helps refine models, improve interpretability, and optimize computational efficiency. Model-agnostic SHapley Additive exPlanations (SHAP) values are frequently used (=21) [21, 39, 86, 88, 91, 92, 95, 100, 102, 104, 111, 114, 115, 123–125, 131–133, 135, 137] to interpret the contribution of different features, offering transparency in model decision-making.

Summary. Many reviewed studies apply EDA using dimensionality reduction techniques such as PCA and t-SNE, as well as clustering methods including UHCA and dendrograms, primarily to visualize class separability or explore spectral relationships. Feature importance analysis, particularly through SHAP values, is increasingly adopted to enhance interpretability. However, EDA remains inconsistently applied and often superficially interpreted across studies. Given the complexity of MALDI-TOF MS data—including high dimensionality, variability, missing values, and potential outliers—EDA should play a more central role within ML pipelines. Thorough analysis can reveal anomalies, assess distributional properties, guide imputation strategies, and inform critical downstream decisions such as preprocessing, feature selection, and feature engineering. Moreover, systematic EDA contributes directly to model interpretability by enabling more informed ML decisions, reducing black-box behavior, and supporting explainable modeling strategies. This aspect is particularly critical in medical applications, where transparency, reliability, and clinical applicability strongly depend on understanding both data characteristics and model behavior. A more structured and systematically reported use of EDA would therefore improve data comprehension, reduce modeling biases, and enhance the robustness and reliability of ML-based findings.

Preprocessing

Preprocessing is a critical step in MALDI-TOF MS data analysis, directly influencing ML model performance. The principle of ”garbage in, garbage out” [152] strongly applies here, emphasizing the need for standardized workflows to ensure high-quality data input. Spectrometry data typically contain noise, such as baseline shifts and fluctuations, which require correction through baseline subtraction and smoothing techniques. Additionally, not all regions of the mass spectrum are informative; peak selection (PS) methods help isolate relevant signals (see section Feature selection). Normalization is essential due to the wide intensity range in spectra, often spanning thousands of units. Furthermore, spectra acquired from different instruments (see section Data description) exhibit variability, necessitating standardization to ensure robust, generalizable models. Since MALDI-TOF MS data can vary in length, binning techniques are often applied to create fixed-size inputs for ML models. Lastly, technical replicates—multiple spectra from the same sample—must be carefully handled to prevent duplicated data and potential bias.

Two widely used preprocessing tools are the open-source R package MALDIquant [138] (=23, see Table 1), and ClinProTools [153] (=21) [20, 28, 30, 42, 44–51, 53, 56, 64, 67, 71, 72, 74, 109, 130], a commercial software integrated into Bruker Daltonics’ [143] systems. Both offer extensive preprocessing options, but the lack of standardization in parameter selection and processing order presents a major challenge, hindering reproducibility and comparability across studies. This variability results in diverse preprocessing pipelines, e.g. Weis [25] applied normalization, Savitzky–Golay smoothing, baseline correction, Total Ion Current (TIC) normalization, signal-to-noise ratio (SNR) thresholding, warping, and trimming, whereas Dematheis [80] employed spectral cleaning, transformation, smoothing, baseline correction, and normalization.

Despite methodological discrepancies, several preprocessing steps are consistently applied across studies, including baseline subtraction (=59), smoothing (=51; e.g. Savitzky–Golay), normalization (=52; e.g. TIC), SNR thresholding (=26), spectra averaging (=24), and binning (=44). Additional techniques, such as scaling, spectra, and peak alignment, recalibration, binarization, and transformations (e.g. wavelet or helix matrix), are less commonly reported. Table 1 provides a comprehensive overview of the principal steps applied and whether a study includes ablation or comparative studies, indicated by an asterisk (*).

Importantly, only a limited number of studies systematically evaluate how preprocessing choices influence downstream performance. A notable exception is the MSclassifR study by Godmer et al. [121], which performed a large-scale comparative and ablation analysis of preprocessing strategies. Their results indicate that individual choices for smoothing (Wavelet versus Savitzky–Golay), baseline correction (SNIP versus TopHat), and normalization (TIC versus PQN) often yield statistically comparable performance. In contrast, spectral alignment emerged as a decisive factor, with LOWESS-based alignment significantly outperforming alternative approaches. Furthermore, while intensity transformations showed minor differences, the overall pipeline performance proved more robust than any isolated step. This study highlights several practical takeaways: (i) baseline correction, normalization, and alignment are non-negotiable steps; (ii) alignment quality can dominate performance variations; and (iii) excessive focus on marginally different smoothing or calibration methods may offer limited gains relative to pipeline consistency. Readers seeking a deeper understanding of preprocessing effects are encouraged to consult this work directly [121].

Summary. The lack of standardization in MALDI-TOF MS preprocessing impacts data quality, model performance, and reproducibility. Studies vary widely in method selection, ordering, and hyperparameter choices, often without clear justification, making replication and optimization difficult. While tools like MALDIquant and ClinProTools provide automated workflows, their customization options introduce inconsistencies that hinder comparability. Comparative evidence indicates that several widely used techniques are largely interchangeable in terms of performance, while spectral alignment may represent a critical determinant of predictive stability. Consequently, baseline correction, normalization, and alignment emerge as essential steps, whereas fine-grained differences between smoothing or transformation strategies appear to play a secondary role relative to pipeline coherence.

Feature selection

Feature selection, or PS in this context, is a crucial step in handling high-dimensional datasets. In fact, most reviewed studies (=82) employ PS techniques to enhance model interpretability, improve computational efficiency, and mitigate overfitting. The most common approach involves applying cutoffs, either based on SNR (=25) or intensity thresholds (=3) to eliminate minimally informative variables. Additionally, some studies use statistical methods (=12), such as Pearson Correlation Coefficient, t-tests, or analysis of variance, while AUC-based selection (=4) prioritizes features with the highest discriminative power.

ML-based PS is also widely adopted, with RF-based importance ranking being the most frequently applied technique (=23). Genetic algorithms (GAs) (=8) are explored as well, leveraging evolutionary strategies to optimize feature subsets. Moreover, several studies (=29) utilize alternative PS methods, such as LASSO (=3) and Bayesian automatic relevance determination analysis (=1). For a comprehensive overview of the PS techniques used in each study, please refer to Table 1.

Summary. PS plays a central role in MALDI-TOF MS-based ML pipelines, yet the reviewed literature reveals substantial methodological heterogeneity. PS strategies differ fundamentally in how features are evaluated and should therefore be interpreted with caution. Filter methods (e.g. statistical tests, SNR, or intensity thresholds), while computationally efficient and easy to implement, assess peaks independently and ignore inter-peak relationships, potentially overlooking biologically meaningful spectral patterns. Wrapper (e.g. forward/backward selection, recursive feature elimination) and embedded (e.g. LASSO, RF importance) methods account for feature dependencies but may reduce interpretability or introduce instability, particularly in small datasets. Consequently, PS design involves an inherent trade-off between performance, interpretability, and biological plausibility—a balance that remains insufficiently discussed. Importantly, emerging modeling strategies challenge the necessity of predefined PS. Approaches operating on variable-length spectra or sparse representations, including transformer-based models, may reduce information loss and better capture spectral structure. Exploring ML pipelines without PS, or combining PS with representation learning techniques, represents a promising direction for improving robustness and clinical applicability.

Algorithms

A wide range of algorithms (=34) is applied across the reviewed studies, with most (=83) comparing multiple models to assess performance. However, algorithm selection is generally driven by dataset size, quality, and computational constraints rather than the specific task (ID, AMR, or TYP), with minimal correlation between model choice and study objective.

Figure 3 illustrates the evolution of best-performing algorithms over time. The increasing use of ensemble models, such as RF and Boosting, along with the emergence of DL approaches—particularly after the release of the DRIAMS database—reflects a shift toward more complex, data-driven methodologies. Despite this, clinical applications often prioritize interpretability. As a result, classical ML models such as RF (=37) and SVM (=22) remain the most widely used, owing to their robustness and ability to handle high-dimensional data. RF is particularly favored for its inherent interpretability, which stems from its ensemble structure and feature importance measures. SVMs also offer interpretability when linear kernels are used; however, this advantage diminishes with more complex kernels. Nevertheless, SVMs continue to be popular, especially in settings where the number of features exceeds the number of samples [73, 75, 77, 79, 83, 85, 114, 115].

Boosting methods (=20), including GB, AdaB [154], LGBM, and XGB [155], also demonstrate strong performance. Less common techniques include GA (=10), KNN (=5), discriminant analysis (PLS-DA, LDA) (=3), and probabilistic approaches (Gaussian processes, Factor Analysis) (=3). Overall, classical ML models outnumber DL approaches (=23)—such as Multi-Layer Perceptrons (MLP), Convolutional Neural Networks, Autoencoders (AE), Long Short-Term Memory (LSTM), and Transformer models. DL models are advantageous when applied to raw spectra, as they excel at extracting local patterns from mass spectrometry data. However, their effectiveness depends on large datasets [15, 39, 65, 78], with classical ML models often outperforming them when data are limited [118, 121, 124]. Additionally, =2 studies use two-step models that combine different algorithms to enhance classification [16, 30]. For example [16], applied a Denoising AE for dimensionality reduction, feeding its latent space into an SVM.

Only =2 studies [114, 120] employ transformer-based architectures. Nguyen et al. [114] apply transformers within a transfer learning framework, where a pretrained bacterial identification model is fine-tuned for AMR prediction using a meta-learning strategy. Their approach subsequently integrates five base classifiers (LR, RF, SVM, LGBM, and MLP), leveraging transformer-derived latent embeddings to achieve the highest predictive performance.

More recently, De Waele et al. [120] introduce a transformer architecture specifically designed for MALDI-TOF MS data, combined with a self-supervised pre-training strategy. Their results demonstrate that transformer-based representations achieve state-of-the-art (SOTA) or competitive performance across multiple downstream tasks, including species ID and AMR prediction. Notably, the benefits of this transformer are most pronounced in complex classification settings, such as large-scale multi-class identification and AMR prediction, while performance gains are less evident in simpler tasks. Furthermore, the authors emphasize the critical role of pretraining, showing that transformers trained from scratch exhibit inferior performance, whereas self-supervised pretraining substantially improves generalization.

Summary. The reviewed literature highlights a critical methodological observation: increased model complexity does not inherently guarantee improved predictive performance. Classical ML approaches, particularly RF and boosting-based models, consistently demonstrate strong and stable results across a wide range of MALDI-TOF MS applications, indicating high data efficiency in typical small-sample, high-dimensional settings.

Recent studies [120] provide important nuance to this interpretation. While transformer architectures achieve superior performance across multiple downstream tasks, their advantages appear strongly task-dependent. In AMR prediction, transformer-derived representations consistently outperform MLP baselines in AUROC, suggesting improved spectral embedding quality. However, for simpler species ID scenarios, performance gains are less pronounced, with classical methods remaining competitive. Notably, experimental analyses show that the performance advantage of transformer models diminishes as task complexity decreases, reinforcing that architectural sophistication primarily benefits complex classification problems.

Collectively, these findings caution against equating methodological complexity with performance improvement. Instead, coherent ML model design requires careful alignment between task difficulty, dataset characteristics, and modeling strategy. Meaningful comparisons between algorithms are only valid under comparable experimental conditions. Consequently, selecting models based on data regime and problem structure, rather than methodological novelty alone, emerges as a central principle for developing robust and clinically relevant MALDI-TOF MS-based ML systems.

Evaluation metrics

The reviewed studies employ various evaluation metrics to assess model performance. Accuracy is the most frequently reported metric (=92) [14, 17–20, 22, 23, 28–31, 40, 41, 43–57, 59, 61–65, 71–73, 81–87, 90–104, 106–109, 111, 113, 116–118, 120–124, 128, 129, 133–137], providing a general measure of correctly classified instances. However, given the potential class imbalances in AMR prediction and bacterial ID and TYP tasks, relying solely on accuracy can be misleading. To address this, sensitivity (recall) (=70) [16, 17, 19–22, 27–31, 42, 45, 46, 48–51, 56, 57, 62, 64–68, 71–78, 80, 82–87, 90, 92–100, 102–104, 108, 109, 113, 114, 116, 118, 119, 123, 125, 127–131, 133, 136] and specificity (=64) [16, 17, 19–22, 27–31, 42, 45, 46, 48–51, 56, 57, 62, 64–68, 71–78, 80, 82–87, 90, 92–100, 102–104, 108, 109, 113, 114, 116, 118, 119, 123–125, 127, 128, 130, 131, 133, 136] are commonly reported, measuring the model’s ability to correctly identify positive and negative cases, respectively. Positive/negative predictive value (=27)[18, 23, 27, 29, 31, 39, 42, 46, 70, 74, 76, 83, 87, 93, 94, 106–108, 112, 117–119, 125, 127, 131, 134, 136]/(=22[23, 27, 29, 31, 39, 42, 46, 70, 74, 76, 83, 87, 93, 107, 108, 112, 118, 119, 125, 127, 131, 136]) assesses the reliability of predictions, while precision (=8)[18, 31, 65, 84, 98, 116, 120, 129], which quantifies the proportion of correctly identified positives out of all predicted positives, is less frequently used but remains valuable when false positives are costly. These metrics are particularly critical in AMR tasks, where misclassifying resistant bacteria as susceptible (false negatives) can have serious clinical implications.

F1-score (=25), which balances precision and recall, is especially useful for imbalanced datasets [16, 23, 65, 66, 76, 83, 87, 91, 92, 102, 104, 107, 108, 111, 112, 114, 118, 119, 124, 127–129, 131, 133, 137]. AUROC (=66) is also widely used [15, 17, 21–23, 29, 30, 39, 55–57, 60–62, 65, 68, 70, 73, 75–79, 82–92, 94–100, 102–108, 110, 112, 114, 115, 117–120, 123–125, 127–129, 131, 132, 135, 137], offering an aggregate measure of a model’s discriminative ability across different thresholds. To further address class imbalance, the AUPRC (=17) has gained popularity [15, 21, 25, 29, 39, 76, 88, 91, 102, 104, 114, 115, 125, 128, 132, 133, 137], as it provides a more reliable assessment when the minority class is of primary interest (see section Data description).

Less frequently reported metrics include confusion matrix analysis (=6) [18, 22, 59, 66, 80, 99], Kappa (=8) [31, 69, 82, 93, 118, 121, 122, 125], Youden’s index (=3) [46, 60, 100], Matthews correlation coefficient (MCC) (=4) [64, 68, 73, 128], mean error (ME) (=2) [39, 73], Jaccard index (=1) [84], Hamming loss (=1) [111], and G-measure (=1) [113].

Summary. While traditional metrics such as accuracy, sensitivity, and specificity remain dominant, the increasing adoption of AUROC, AUPRC, and F1-score reflects a growing awareness of the challenges posed by imbalanced datasets. The clinical utility of ML performance metrics depends on the specific diagnostic task and its associated risk. In bacterial ID, clinical applicability is often defined by the model’s ability to provide a high-confidence score that distinguishes between species with overlapping spectral profiles. For TYP, the real-world value lies in the sensitivity of the ML model to detect subtle sub-species variations in the proteome that signal outbreaks or hypervirulent lineages. In the case of AMR prediction, spectral classifiers must be evaluated against established regulatory safety benchmarks; for instance, the ISO 20776-2 and FDA standards for AST require that Very Major Errors (VME) remain below 1.5, as failing to detect a resistant proteomic phenotype can lead to treatment failure [156, 157]. Across all these tasks, moving beyond aggregate metrics like accuracy to class-specific performance is essential for ensuring that ML-driven spectral analysis is safe for routine clinical use.

Model proceedings

Hyperparameter tuning

Hyperparameter tuning is crucial for ML model optimization, directly influencing predictive performance and generalizability. However, the extent of tuning varies significantly across the reviewed studies. A substantial number (=25) do not disclose any details on hyperparameter selection [16, 27, 30, 42, 43, 45, 46, 49, 52–54, 57–59, 61–64, 69, 72, 92, 93, 100, 126, 130], leaving it unclear whether parameters were optimized or left at default values. Additionally, =40 rely entirely on default settings [18–20, 28, 29, 41, 44, 47, 48, 50, 51, 65–67, 70, 71, 78, 79, 81, 83, 85, 87, 94–96, 98, 99, 102, 104, 106, 116, 117, 122, 123, 125, 128, 129, 133, 135, 136].

Among studies that perform hyperparameter optimization, CV is the most common approach (=40) [15, 17, 21–23, 25, 31, 39, 55, 60, 68, 73–77, 80, 84, 86, 88–91, 97, 101, 105, 107–110, 112, 113, 115, 119, 121, 124, 127, 131, 132, 134]. The CV method optimizes hyperparameters by creating training and validation partitions within the training set, helping to mitigate overfitting and ensure robust model selection. Only =4 studies apply Bayesian hyperparameter selection [111, 114, 134, 137].

A smaller subset (=4) employs manual ablation studies [14, 40, 56, 103, 118], though this approach is often inefficient and lacks reproducibility.

Evaluation strategy

Evaluating model performance is a crucial step in ML workflows, ensuring that predictive models generalize well beyond their training data. The reviewed studies employ a variety of evaluation strategies, with two studies not specifying their approach (=2) [61, 126].

Predefined train-test splits are used in =64 studies [15, 21–23, 27, 31, 41, 46, 49, 53, 63, 65, 76, 77, 82, 84, 86, 97, 100, 105, 107–109, 111–113, 115, 116, 119–125, 127, 128, 130, 131, 133–136]. In these cases, a fixed portion of the dataset is allocated for training, while the remaining data are reserved for posterior evaluation. This approach is computationally efficient but may introduce bias if the split is not representative of the full data distribution. =30 also apply CV on the predefined training set to improve model robustness [14, 18, 20, 29, 39, 42, 51, 54, 66, 67, 69, 70, 74, 75, 79–81, 89, 99, 102, 110, 116, 121–124, 128, 130, 133, 135].

Outer CV is the most common evaluation method, applied in =70 studies [16, 17, 28, 30, 43–45, 47, 48, 50, 55–58, 60, 64, 68, 71–73, 78, 83, 85, 88, 90–92, 94–96, 98, 103, 104, 106, 114, 117, 118, 129, 132, 137]. This technique repeatedly splits the dataset into training and validation folds, reducing overfitting and providing a more reliable estimate of model performance. Within this group, a subset of studies (=4) employs stratified CV, which ensures class distribution is preserved across folds—a particularly valuable approach for imbalanced datasets [25, 62, 87, 89]. In contrast, =7 studies rely on leave-one-out (LOO) CV for model validation [19, 40, 42, 52, 59, 101, 121]. This method iteratively trains the model on all but one sample, testing it on the excluded instance until every sample has been used for validation. While LOO-CV provides an almost unbiased estimate of model performance, it comes at a high computational cost and suffers from increased variance due to strong correlations between training sets. Conversely, lower-fold CV methods reduce variance by training on more distinct subsets but may introduce higher bias when training sets are too small. Given this trade-off, LOO-CV is only justified for very small datasets.

OOD validation, or external validation, refers to the process of evaluating an ML model on an independent dataset that was not used during training or outer CV. This step is crucial for assessing a model’s ability to generalize beyond the data on which it was developed, ensuring its robustness when applied to new or unseen distributions. However, despite its importance, OOD is often overlooked or misunderstood. Many studies conflate OOD validation with standard evaluation methods, such as a predefined test split, that assess performance only within the same dataset. Among the reviewed studies, only =32 actually conduct OOD [15, 17, 21, 39, 42, 58, 64, 68–71, 73, 78, 84, 86, 88, 90, 91, 93, 94, 100, 103, 111, 114, 115, 120, 121, 123, 128, 131, 132, 134]. The remaining studies rely solely on internal validation.

Summary. Model development practices across the reviewed studies reveal several methodological limitations. A substantial proportion of works (=25) do not disclose hyperparameter configurations, preventing reproducibility and hindering objective assessment of model optimization. When hyperparameters are reported, it remains essential to clarify whether values correspond to default settings or systematic tuning procedures, preferably conducted via cross-validation. Without appropriate tuning, models may operate under suboptimal configurations, potentially underestimating achievable performance. Similarly, studies reporting performance metrics derived from single train-test splits or non-averaged evaluations lack evidence of predictive robustness, as variability across data partitions is not assessed. Reliable ML pipelines require performance estimates aggregated across multiple iterations or validation folds to ensure stability and generalizability. Finally, despite the widespread adoption of internal validation strategies, only =32 studies conduct true OOD validation. While outer CV mitigates overfitting within a dataset, it primarily evaluates interpolation rather than robustness to distribution shifts commonly encountered in MALDI-TOF MS applications. Given the variability introduced by laboratory protocols, acquisition devices, and population heterogeneity, models lacking external validation remain vulnerable to performance degradation in real-world clinical settings. This imbalance between internal optimization and external robustness highlights a persistent gap between methodological evaluation and clinical deployment requirements.

Open-source data and code

The availability of open-source data and code is essential for advancing research, promoting transparency, and ensuring reproducibility in ML studies, including sharing software and implementation details. Python is the predominant software (=47) [15, 18, 21–23, 25, 27, 39, 65, 70, 73, 76–78, 84–86, 88–90, 92, 95–97, 100, 102, 103, 105, 107, 108, 110–115, 119, 120, 124, 127–129, 131, 133–135, 137] due to its versatility and extensive ML libraries [155, 160, 161]. Additionally, Python is more commonly associated with recent and complex studies, with of publications in the last three years employing this programming language. Bruker’s ClinProTools [143] (=21) [20, 28, 30, 42, 44–51, 53, 56, 64, 67, 71, 72, 74, 109, 130] leads among manufacturer-specific tools, followed by Bionumerics [162] (=2) [58, 81] and AMRQuest [163] (=2) [82, 125]. R [164] (=22) [14, 31, 52, 59, 60, 62, 63, 66, 68, 70, 75, 79, 80, 93, 98, 116, 118, 121, 122, 126, 132, 136] is favored for its robust statistical analysis capabilities and specialized mass spectrometry tools [138]. Web-based tools like Clover [165] (=10) [17, 29, 69, 83, 87, 91, 94, 101, 106, 117] offer streamlined clinical workflows but are less flexible. MATLAB [166] (=6) [40, 43, 55, 57, 61, 99] is used for its matrix handling and visualization, though its closed-source nature limits adoption. Only =4 studies omit software details, hindering reproducibility[16, 41, 104, 123].

Among the reviewed studies, only a small subset publicly shares their datasets (=15) and (=17) share their ML implementation, which is crucial for full reproducibility (see Table 4). Despite the growing availability of MALDI-TOF MS datasets, only two contain data labeled for TYP [93, 96]. The absence of open-access typed data restricts researchers’ ability to develop and validate models for hypervirulence prediction across bacterial strains.

Table 4.

Summary of studies providing open-source datasets and/or code, listing dataset availability, sizes, formats (raw or preprocessed, pp) code accessibility, software used, and corresponding repository links.

Paper	Data	Size	Format	Code	Software	Links
Asakura [59]	✓	N/A	pp	✓	C++/R	GitHub
Desaire [14]	✓	571	raw			UCI
Ling [65]				✓	Python	GitHub
Weis [25]				✓	Python	GitHub
Weis [21]	✓	132 097	raw, pp	✓	Python	DRYAD, GitHub
Wang [78]				✓	Python	GitHub
Yu [86]	✓	2456	pp			Zenodo
Wang [88]	✓	31 807	pp	✓	R	GitHub
G.-López [89]	✓	2000	raw, pp	✓	Python	GitHub
Candela [17]	✓	1566	raw			Clover
Chung [90]	✓	37 918	pp			GitHub
Yu [92]	✓	1853	pp			Zenodo
Zintgraff [93]	✓	193	raw	✓	R	Mendeley
Wang [96]	✓	205	pp	✓	Python	Article
de Waele [105]				✓	Python	GitHub
Mao [18]	✓	N/A	pp			SimTK
L.-Cortés [39]				✓	Python	GitHub
Astudillo [111]				✓	Python	GitHub
Nguyen [114]	✓	30	raw	✓	Python	GitHub
Ren [115]				✓	Python	GitHub
de Waele [120]				✓	Python	Github
Godmer [121]				✓	R	Github
B.-Sánchez [134]	✓	379	raw	✓	Python	Zenodo, Github
Thai [136]	✓	50	raw			Mendeley
RKI dataset [158]	✓	11 038	raw			Zenodo
MARISMa [36]	✓	202 700	raw			Zenodo
MS-UMG [159]	✓	77 342	pp			Zenodo
Sum:	18			17

Among the open datasets, file formats vary considerably, with few studies (=2) providing raw spectral data (e.g. BrukerRaw format). In contrast, others release raw or preprocessed data in the non-original format (csv, txt, npy), limiting its usability for other research purposes. As shown in Table 4, only a few datasets retain the original metadata, such as acquisition time or device information, which is crucial for standardization and reproducibility.

Besides the DRIAMS [35] dataset published by [21], other large public MALDI-TOF MS datasets are available to the research community: the RKI [158], MARISMa [36], and MS-UMG [159] datasets. All provide a comprehensive collection of spectra from various bacterial species, enabling large-scale studies on bacterial ID and AMR (except RKI). These repositories have facilitated replicating and validating existing models and spurred new research. A total of =9 studies use the DRIAMS and/or RKI datasets for their research [39, 97, 104, 105, 111, 113, 115, 120, 132], illustrating the transformative impact of open-access data on advancing the field.

Summary. Open-science practices play a fundamental role in ensuring reliability, transparency, and reproducibility in ML-driven microbiology research. Publicly accessible datasets and well-documented code enable independent validation of reported findings, facilitate methodological comparisons, and allow other laboratories to assess model robustness on their own data. In clinical contexts, where variability across instruments, protocols, and populations is unavoidable, reproducibility becomes a prerequisite for trustworthy model deployment. Beyond reproducibility, open data sharing has the potential to substantially accelerate methodological innovation. Large-scale, heterogeneous MALDI-TOF MS repositories incorporating diverse isolates, strain-level variability, resistance phenotypes, and epidemiological heterogeneity would enable the development of more generalizable models and foster emerging research directions, including representation learning, generative AI, and foundational spectral models. Importantly, aggregating data across laboratories would enhance both biological and technical variability, a critical factor for designing ML systems robust to distribution shifts inherent to real-world clinical environments.

Challenges and future directions

Progress in ML for microbiology critically depends on data and model transparency. The current literature exhibits substantial heterogeneity and incomplete reporting regarding dataset composition, preprocessing pipelines, and model configurations, limiting reproducibility and independent validation. True scientific progress requires clear documentation of data distributions, replicate handling, preprocessing strategies, and hyperparameter settings. Furthermore, emerging microbiological challenges, including closely related species complexes and high-burden resistant pathogens [151], highlight potential gaps between research trends and clinical needs. While certain well-studied problems continue to dominate the literature, other clinically relevant pathogens remain comparatively underexplored. Aligning methodological research with real-world clinical challenges, supported by standardized, large-scale MALDI-TOF MS repositories, represents an essential direction for future work and could substantially accelerate methodological development while improving clinical relevance [167].

Another major challenge lies in adequately characterizing variability inherent to MALDI-TOF MS data. Spectral variability arising from laboratory protocols, acquisition devices, culture conditions, and temporal drift can substantially influence model behavior. Consequently, thorough exploratory data analysis prior to modeling becomes essential, incorporating both interpretability and variability assessment. Recent studies demonstrate that variability-aware data augmentation strategies can mitigate performance degradation and improve model robustness under heterogeneous acquisition conditions [33]. In parallel, domain adaptation and domain alignment methods represent promising directions for addressing distribution shifts, yet their adoption within MALDI-TOF MS workflows remains unexplored.

Consistency in performance assessment also represents a critical methodological requirement. The reviewed studies reveal substantial variability in reporting practices, including incomplete disclosure of hyperparameter tuning procedures, inconsistent feature selection strategies, and reliance on non-averaged evaluation metrics. Proper ML pipelines require systematic hyperparameter optimization tailored to the specific data environment, alongside rigorous validation strategies. Performance estimates should be aggregated across multiple data folds to ensure robustness. Importantly, clinical ML applications must account for error asymmetry, such as in AMR prediction, where false-negative errors may have disproportionately severe clinical consequences. Consequently, evaluation metrics sensitive to minority-class performance and clinical risk, such as sensitivity, recall, AUPRC, or balanced accuracy, should be prioritized.

Finally, external validation remains one of the most significant barriers to clinical applicability. While internal validation strategies are widely adopted, only a limited number of studies perform true OOD evaluation. Given the variability and distribution shifts inherent to clinical microbiology data, models validated exclusively within a single dataset may exhibit reduced transportability across laboratories or clinical environments. Robust OOD validation is therefore essential for assessing model reliability, stability, and real-world clinical utility [168].

Recent advances in DL and representation learning offer promising directions for MALDI-TOF MS analysis, although DL methods remain limited in the reviewed literature. Observed performance trends indicate that for simpler tasks, such as binary AMR prediction or restricted multi-class classification, increased model complexity does not consistently yield performance gains. Classical ML approaches, particularly RF and boosting-based models, often provide robust and competitive results, highlighting their strong data efficiency. In contrast, representation-learning strategies, including multivariate methods such as PLS and transformer architectures leveraging self-supervised pretraining, demonstrate clearer advantages in more complex scenarios by effectively capturing spectral variability and improving generalization. Complementing these developments, recent generative AI approaches enabling conditional spectrum synthesis [169] present a potential solution to data scarcity and class imbalance, offering new opportunities for data augmentation and variability-aware model development.

Guidelines

After reviewing =115 SOTA papers, we identified a lack of standardization in ML research on MALDI-TOF MS data. To address this, we propose a set of guidelines to establish a well-structured research pipeline that can effectively advance ML applications in this field. A clear definition of the research task is crucial, specifying how data will be utilized and how ML techniques will be applied. Once this is defined, the pipeline should follow key stages: (i) EDA, (ii) spectra preprocessing and PS, (iii) model selection, (iv) training, and (v) evaluation, as in Fig. 4.

Figure 4.

Overview of key steps in ML-based MALDI-TOF MS analysis, covering exploratory data analysis (EDA), preprocessing, model selection, training, and evaluation, highlighting considerations such as data imbalance handling, hyperparameter tuning, and cross-validation (CV), with recommended steps in black, alternative approaches in blue, and optional steps in green.

During the EDA, a comprehensive dataset analysis is crucial for making informed preprocessing and model selection decisions. One of the first aspects to consider is the number of unique bacterial isolates, as this directly influences model selection. Additionally, handling spectral replicates is essential, and there are two primary approaches: (1) either peak alignment and averaging spectra to create a single representation per isolate or (2) treating them as a multi-instance problem. The latter approach is preferable, as it maximizes data usage and improves reproducibility, but care must be taken to ensure that replicates remain within the same split in both inner and outer CV to prevent data leakage. Another key factor is identifying data imbalance, which has a direct impact on loss functions, and evaluation strategies. Finally, data visualization using unsupervised learning techniques, e.g. PCA or UHCA, can enhance interpretability and pattern recognition. However, these methods should be reserved for exploratory analysis rather than classification tasks, which require supervised learning models.

Preprocessing plays a fundamental role in standardizing datasets for better model learning. Although current SOTA practices lack consistency, we recommend several essential steps for raw spectral data: (i) variance stabilization using log2 or square root transformation, (ii) smoothing with a Savitzky–Golay filter, (iii) baseline removal with techniques such as the Statistics-sensitive Non-linear Iterative Peak-clipping algorithm (SNIP) or top-hat transformation, (iv) normalization of the intensities, and (v) trimming within the range of 2000 to 20 000 Da to eliminate noise. After this, spectra can either be used in their original varying lengths or standardized. A straightforward approach is binning with a 3 Da bin size, converting spectra into 6000-element vectors by summing intensities within each bin. Alternatively, PS techniques, e.g. based on RF, can be employed to retain the most relevant features. For more advanced implementations, models such as Transformers [114] can directly handle variable-length spectra.

Choosing the right model depends on the research question and the characteristics of the dataset. For small- to medium-sized datasets, classical ML models such as RF and LGBM generally provide strong performance. Bayesian models, such as GPs [25] or FA [89], are particularly useful for combining multimodal data, as they generalize well and do not require an inner CV. In contrast, larger datasets allow for the use of DL models [21, 114], which are highly effective at capturing complex representations. Before model training, data imbalance must be addressed, which can be done through either (1) oversampling techniques like SMOTE [113, 114] or more targeted approaches [33], or by (2) incorporating imbalance-aware loss functions, such as weighted Log-Loss for logistic regression, weighted Hinge Loss for SVMs, or class-weight adjustments in RF models. Hyperparameter tuning is also a key step and should be performed using inner CV, calculating the mean and standard deviation of evaluation metrics to identify the most robust model. As in previous steps, spectral replicates must remain within the same split to prevent bias.

Model evaluation should follow training using outer CV, where the dataset is split into folds (typically 3, 5, or 10), allowing the model to be trained and tested times. Once the best-performing model is identified, it should be retrained using the entire internal dataset. To assess generalizability, an OOD evaluation should be performed using a completely independent dataset, ideally from a different machine or hospital. Evaluation metrics should address class imbalance, with balanced accuracy as a priority, complemented by AUROC, precision, recall, and AUPRC for a comprehensive performance assessment.

Conclusions

The integration of ML into MALDI-TOF MS analysis has significantly advanced bacterial ID, AMR prediction, and strain TYP. However, several challenges remain, including data scarcity, insufficient transparency in reporting methodological details, non-standardized preprocessing workflows, inconsistencies in model evaluation strategies, and limited availability of open-source datasets and code. While classical ML models remain the predominant and often highly competitive approach, recent evidence suggests that increased model complexity does not inherently guarantee superior performance. Instead, performance gains appear strongly dependent on task complexity, dataset characteristics, and the ability to learn robust spectral representations. In this context, representation learning strategies, including transfer learning and self-supervised pretraining, emerge as key drivers of improvement, particularly for complex prediction scenarios. Despite these methodological advances, issues related to reproducibility, robustness, and generalizability continue to hinder reliable clinical deployment.

We hope this review not only provides a comprehensive overview of the current landscape but also serves as a guide toward more standardized and methodologically coherent practices in ML-based MALDI-TOF MS research. By promoting open science, transparent reporting, and rigorous external validation, we aim to support the development of robust, interpretable, and clinically relevant models. Furthermore, we seek to inspire future efforts to better align modeling strategies with clinical realities, foster interdisciplinary collaboration between data scientists and clinical experts, and ultimately drive broader clinical adoption of ML-enhanced MALDI-TOF MS technologies.

Key Points

• This review fills a crucial gap by extensively analyzing machine learning pipeline quality criteria in MALDI-TOF MS applications, a perspective not addressed in existing reviews.

• We find a widespread lack of standardization across key steps, such as preprocessing, peak selection, and model training, many studies also failing in clearly reporting these methodologies.

• Despite the potential of deep learning, it remains underutilized in the field, with most works relying on traditional machine learning methods, thereby limiting MALDI-TOF MS data use for complex tasks in clinical decision-making.

• Research progress is hindered by the limited availability of large, open, and collaborative datasets, which also affects reproducibility, robustness, and generalization of developed models.

• To support the community, we provide actionable guidelines for applying machine learning in MALDI-TOF MS studies, aiming to promote more consistent and clinically useful research practices.

Conflicts of interest

None declared.

Funding

This work was supported by the Ministry of Science, Innovation and Universities—Agencia Estatal de Investigación (MCIN/AEI/10.13039/501100011033/FEDER, UE) [grant number PID2023-146684NB-I00 to V.G.V.]; and the Community of Madrid [project TEC-2024/COM-89 to V.G.V.]. A.G.L. is supported by the European Union’s Horizon Europe programme through a Marie Skłodowska-Curie Actions Postdoctoral Fellowship (OUTBRAID, Grand Agreement No. 101278403). The work of C.S.S. was supported in part by the Comunidad de Madrid’s 2025 Cesar Nombela program under Grant 2025-T1/COM-36091.

Data availability

This article is a systematic review of previously published studies. No new datasets were generated or analyzed during the current study.