Multimodal deep learning for predicting neoadjuvant treatment outcomes in breast cancer: a systematic review

Abstract

Background

Pathological complete response (pCR) to neoadjuvant systemic therapy (NAST) is an established prognostic marker in breast cancer (BC). Multimodal deep learning (DL), integrating diverse data sources (radiology, pathology, omics, clinical), holds promise for improving pCR prediction accuracy. This systematic review synthesizes evidence on multimodal DL for pCR prediction and compares its performance against unimodal DL.

Methods

Following PRISMA, we searched PubMed, Embase, and Web of Science (January 2015–April 2025) for studies applying DL to predict pCR in BC patients receiving NAST, using data from radiology, digital pathology (DP), multi-omics, and/or clinical records, and reporting AUC. Data on study design, DL architectures, and performance (AUC) were extracted. A narrative synthesis was conducted due to heterogeneity.

Results

Fifty-one studies, mostly retrospective (90.2%, median cohort 281), were included. Magnetic resonance imaging and DP were common primary modalities. Multimodal approaches were used in 52.9% of studies, often combining imaging with clinical data. Convolutional neural networks were the dominant architecture (88.2%). Longitudinal imaging improved prediction over baseline-only (median AUC 0.91 vs. 0.82). Overall, the median AUC across studies was 0.88, with 35.3% achieving AUC ≥ 0.90. Multimodal models showed a modest but consistent improvement over unimodal approaches (median AUC 0.88 vs. 0.83). Omics and clinical text were rarely primary DL inputs.

Conclusion

DL models demonstrate promising accuracy for pCR prediction, especially when integrating multiple modalities and longitudinal imaging. However, significant methodological heterogeneity, reliance on retrospective data, and limited external validation hinder clinical translation. Future research should prioritize prospective validation, integration underutilized data (multi-omics, clinical), and explainable AI to advance DL predictors to the clinical setting.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13062-025-00661-8.

Introduction

Breast cancer (BC) is a heterogeneous disease where neoadjuvant systemic therapy (NAST) may improve surgical and survival outcomes [1]. The achievement of pathological complete response (pCR), defined as the absence of invasive tumor cells in the breast and axillary lymph nodes (ypT0/is, ypN0) [2], is a key prognostic indicator. Several studies, including meta-analyses, have demonstrated that patients achieving pCR often experience significantly better disease-free survival (DFS) and overall survival (OS), particularly in aggressive subtypes such as triple-negative (TNBC) and HER2-positive BC (HER2+ BC) [3, 4]. Given its strong association with long-term outcomes, pCR prediction has become clinically important for tailoring personalized treatment strategies [5]. Early prediction models primarily incorporated clinical variables (tumor size, stage, hormone receptor status, HER2 status) into classical statistical methods like logistic regression-based nomograms [6, 7]. With the emergence of radiomics, handcrafted imaging features extracted from magnetic resonance imaging (MRI) or mammography (MG) were combined with classical machine learning (CML) algorithms (e.g., support vector machine, random forest) to enhance pCR prediction [8, 9]. Parallel to radiomics, quantitative pathomics in digital pathology (DP) imaging and molecular multi-omics expanded the opportunities for predictive modeling, though these often relied on manual or semi-automatic feature engineering [10, 11].

Deep learning (DL)—defined as a type of machine learning (ML) based on artificial neural networks in which multiple layers of processing are used to extract progressively higher level features from data—is an important development in the context of artificial intelligence (AI) technologies [12]. DL architectures, particularly convolutional neural networks (CNNs), have demonstrated significant performance gains in image-related tasks across various medical fields [13]. Their capacity to automatically extract hierarchical features from imaging data has facilitated investigations into predicting pCR in BC without the need for handcrafted feature engineering [14]. Other DL approaches, including autoencoders (AEs), have demonstrated significant potential for extracting, processing, and integrating deep features from high-dimensional molecular tumor profiling data, such as genomic sequencing, epigenomics, transcriptomics, and proteomics [15, 16]. Such frameworks have demonstrated effectiveness in classification tasks for cancer management [17]. Advanced DL models, including large language models (LLMs) or smaller domain-specific language models (e.g. BioBERT), have similarly demonstrated effectiveness in extracting latent representations from biomedical text data [18]. The transition towards multimodal DL, which combines imaging (radiological and histological), molecular profiling, and clinical data, offers a comprehensive approach to capturing the complexity of tumor biology and patient-specific factors. Multimodal DL models integrating clinical data with pretreatment imaging have been developed to predict pCR to neoadjuvant chemotherapy in BC, demonstrating improved predictive performance [19]. Cross-modal approaches that combine DP imaging with longitudinal radiological imaging data acquired during treatment have also shown significant potential for early and accurate prediction of pCR [20]. Recent literature highlights the potential of multimodal DL approaches in precision oncology, emphasizing the need for further research to fully harness their clinical potential [21].

Despite the expanding body of literature, a comprehensive synthesis of studies focusing on DL-driven, multimodal pCR predictors in BC remains limited. Thus, this systematic review aims to address this knowledge gap by systematically evaluating current evidence on multimodal DL approaches for predicting pCR in BC patients treated with NAST. We also examine and compare single versus multimodal approaches and summarize methodological aspects, emphasizing strengths, limitations, and opportunities for advancement, particularly through the incorporation of advanced architectures such as transformers and LLMs, prospective multi-center designs, and underexplored data modalities including molecular multi-omics and free-text clinical notes.

Methods

This systematic review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. Below, we detail the search strategy, eligibility criteria, study selection, data extraction, and data synthesis methods.

Search strategy

A comprehensive literature search was conducted in PubMed/MEDLINE, Embase, and Web of Science databases on January 17, 2025, and updated on April 15, 2025. Articles published from January 1, 2015 onwards were considered, restricted to English-language studies involving Humans. The search aimed to identify investigations evaluating AI methodologies applied to neoadjuvant treatment in BC, using a structured combination of controlled vocabulary and free-text keywords categorized into three thematic areas: 1) the BC domain, including terms such as "breast cancer", "breast carcinoma" and "breast neoplasm"; 2) the NAST domain, including keywords encompassed "neoadjuvant therapy", "neoadjuvant chemotherapy", "neoadjuvant immunotherapy" and "preoperative therapy"; and 3) the the AI domain covering terms including “artificial intelligence”, "machine learning," "deep learning," "neural networks," "transformers," "large language models", “autoencoders”, and associated acronyms (AI, ML, LLM, CNN, GAN, GPT, BERT), as well as advanced techniques like "federated learning", "transfer learning", "ensemble learning", "self-supervised learnin"g, and "multimodal AI". These three domains were combined using the Boolean operator "AND," and the strategy was customized according to the indexing terms and syntax of each database. The complete search strategy and exact syntax used for each database are detailed in the supplementary methods (Supplementary File 1).

Eligibility criteria

Studies selected included invasive BC patients across various subtypes treated with NAST, involving chemotherapy alone or combined with targeted therapies and immunotherapy. The primary focus was on studies employing DL-based AI approaches to predict NAST effectiveness, specifically pCR. DL architectures were required for either feature extraction (including automatic tumor detection) or classifier/predictor construction. Eligible studies incorporated primary data for the prediction from at least one key domain: (i) radiological imaging [MRI, MG, ultrasound imaging (US), computed tomography (CT), or positron emission tomography (PET)]; (ii) DP (histopathological imaging); (iii) multi-omic tumor profiling (including genomics, epigenomics, transcriptomics, and proteomics); or (iv) clinical information from electronic medical records (EMRs) or textual clinical reports (radiology and pathology reports, or routine molecular biomarkers). Studies had to provide quantitative performance metrics (at least the Area Under the Receiver Operating Characteristic Curve—AUC) specifically on pCR. Included were observational studies (retrospective and prospective), clinical trials featuring AI sub-studies, and methodological articles validated in public BC cohorts. Excluded were editorials, letters, conference abstracts, narrative reviews, systematic reviews, meta-analyses, studies with insufficient sample size (< 30 patients), individual case reports, those lacking primary outcome metrics, metrics on response assessments which were not pCR rate, preclinical model studies and studies that applied AI methods which were not identifiable as DL (i.e. neural network with multiple hidden layers).

Study selection and data extraction

All identified studies were managed using Zotero. Following deduplication, six authors (EK, LF, TA, LP, LG, and RK) independently screened titles and abstracts for relevance, ensuring each record was evaluated by at least two authors. Discrepancies were resolved through consensus discussions among these authors, and any persistent disagreements were adjudicated by consultation with senior authors (GB, GC, and PV). Subsequently, selected articles underwent independent full-text review by pairs among the same six authors (EK, LF, TA, LP, LG, and RK), strictly applying predefined inclusion and exclusion criteria. The entire selection and review process, including the number of studies excluded at each stage with detailed reasons, is illustrated in the PRISMA flow diagram (Fig. 1). Data extraction from the final studies was also independently performed by these six authors (EK, LF, TA, LP, LG, and RK) using a structured Microsoft Excel form. Extracted information included study characteristics (authors, publication year, study design, cohort size, data source type), patient demographics, tumor attributes, data modalities utilized (radiological imaging, digital histopathology, multi-omic profiling, clinical data), AI methodologies, and predictive performance metrics.

Fig. 1.

Flow diagram of preferred reporting items for systematic reviews and meta-analyses (PRISMA). DL, deep learning

Data synthesis and analysis

Given the substantial methodological variability observed across studies, encompassing differences in AI approaches such as diverse DL architectures, the distinct types of input data, different multimodal integration strategies, as well as variability in reported performance metrics and outcomes, a rigorous quantitative meta-analysis was not feasible. Consequently, we conducted a comprehensive narrative synthesis. Descriptive summary statistics for all 51 studies, grouped by subtype, study design, primary and secondary modalities, number of imaging time points, and AI technique are provided in supplementary data (Supplementary File 2).

Results

Overview of included studies

A total of 51 studies were selected for final synthesis in this systematic review, all employing DL models to predict pCR in BC patients undergoing NAST. Figure 1 comprehensively illustrates the PRISMA flow diagram, detailing the initial identification of 2,598 records, removal of duplicates (n = 828), exclusion of clearly ineligible records (n = 807), and exclusion based on title and abstract screening (n = 865). Subsequently, 82 full-text articles were assessed for eligibility, and 31 were excluded for specific reasons: non-eligible publication types (n = 3), inadequate/absent DL methodology or data (n = 13), or mismatch in terms of NAST outcome and/or performance metrics (n = 15), ultimately resulting in 51 studies included in this review.

The majority of these studies adopted retrospective designs [n = 46 (90.2%)], with 28 conducted at a single center and the remaining 23 utilizing multi-center cohorts (sometimes relying on public datasets). Sample sizes varied widely, ranging from 31 to 3,352 patients [22, 23], with a median cohort size of 281 patients. Most studies (41/51) enrolled heterogeneous populations comprising multiple BC subtypes [TNBC, HER2+ BC, and hormone receptor-negative/HER2-negative breast cancer (HR+ /HER2- BC)]. Among the remaining studies, five specifically focused on TNBC patients [24–28], two exclusively investigated HER2+ BC [29, 30], one included both HER2+ BC and TNBC patients [31], one targeted patients with HER2-low BC [32], and one examined HER2- BC [33]. Thirty-nine studies (76.5%) validated/tested the final model on an independent patient cohort that was not used during the training phase. This independent cohort could be external (in the case of multi-center studies), internal (a held-out subgroup in single center studies), or a public dataset.

Most of the investigations [n = 37 (72.5%)] leveraged radiological imaging as the primary data modality to predict pCR through DL. MRI was predominant, utilized in 20 studies, followed by US, which served as the main data source in 12 studies. Several other radiological and functional imaging techniques were less commonly investigated, including CT (2 studies) [34, 35], MG (2 studies) [36, 37], and PET (1 study) [22]. In the domain of histopathology, DP was the primary modality in 13 studies, highlighting its relevance in this context. In contrast, only a single study primarily utilized RNA sequencing data [32] for DL analysis (in the domain of molecular -omics profiling), while studies applying DL primarily on textual clinical data to predict pCR were notably absent. Figure 2 provides a Sankey diagram of how each study’s primary modality is combined with secondary and even tertiary inputs. In particular, 27 out of the 51 studies (52.9%) employed multimodal approaches, integrating two or more data types to enhance predictive accuracy. In most cases (24/27), multimodal predictive models combined the primary modality with just one additional data type; clinical data was the most frequently integrated secondary modality, being utilized in 23 studies, while CT was combined with DP in only one study [38]. Among bimodal combinations, the integration of MRI with clinical data was the most common, observed in 11 studies. Conversely, DP frequently appeared as a single modality, being employed alone in 8 of the 13 studies using DP as the primary data source. A smaller subset of three studies explored multimodal integration involving three data modalities. Specifically, two studies utilized US as the primary modality, supplementing it with DP and clinical data [20, 39], whereas the third study employed MRI as the primary modality combined with MG and clinical data [23]. Notably, none of the studies integrated all four data domains (radiological, histopathological, molecular, and clinical) simultaneously to predict pCR, thereby missing potential synergistic gains that multi-modal fusion may offer.

Fig. 2.

Sankey diagram showing the distribution of primary, secondary and tertiary modalities among the 51 studies included in this review. The left column indicates each study’s primary data modality, the middle column its secondary data modality, and the right column its tertiary data modality. Flow widths are proportional to the number of studies integrating a data modality level with the next. MRI, Magnetic Resonance Imaging; DP, Digital pathology; US, Ultrasonography (Ultrasound); CT, Computed tomography; MG, Mammography; PET, Positron emission tomography; RNA, RNA Sequencing data; Clinical: Clinical data; None, No additional data modality. Numbers in parentheses denote study counts (n) for each modality at the given level

With respect to DL strategies, the majority of the studies [n = 45 (88.2%)] utilized CNNs as the core architecture for feature extraction from imaging data. In nine studies, CNNs were combined with advanced DL components, such as attention mechanisms and/or transformer modules (particularly within multimodal frameworks) [36, 37, 40–44], or long short-term memory (LSTM) networks [45, 46] when longitudinal imaging data were incorporated. Five studies employed exclusively transformer architecture (without CNNs) for feature extraction [20, 30, 47–49], with different adaptations of attention mechanisms. Additionally, in five studies, CNNs were supplemented with domain-specific methods, such as handcrafted radiomics or quantitative ultrasound-derived or classical pathomic features [38, 50–53]. Regarding the multimodal integration and/or the classifier, most studies (n = 33) employed end-to-end DL strategies, retaining CNNs or other DL methods throughout the predictive pipeline. Conversely, 18 studies adopted hybrid approaches, limiting DL to the feature extraction step and relying on CML methods, including support vector machines or regularized regression models (e.g., LASSO logistic regression), for integration and final predictions. Longitudinal imaging data acquisition emerged as a relevant factor for pCR prediction. Specifically, 22 studies leveraged imaging data (predominantly MRI and US) collected at multiple time points (ranging from two to five acquisitions at early, mid-, and post-NAST phases) to build predictive models. In contrast, the majority of studies (n = 29) exclusively utilized baseline pre-NAST imaging data.

Reported pCR rates (considering all patients involved in the respective works) varied substantially across studies, ranging from 16.7% [39] to 59.1% [30], with a mean pCR rate of 32.8%. However, this proportion increased to 42.8% when selecting for studies including only patients with TNBC and/or HER2+ BC. All studies reported performance metrics evaluating the ability of the final predictive models to discriminate between pCR and non-pCR cases, most frequently using AUC. Across the reviewed studies, AUC values ranged from 0.70 to 1.00, with a median of 0.88 and a mean of 0.85. Notably, 18 out of the 51 studies (35.3%) reported AUC values equal to or greater than 0.90, highlighting their high discriminative capacity in predicting pCR. Median and mean AUC values improved from 0.83 in the 23 studies using a single data modality to 0.88 and 0.86, respectively, in the 28 studies integrating at least one additional data type. Among the modalities frequently used as standalone data sources, median and mean AUC values were 0.79 and 0.80 for DP (8 studies), 0.84 and 0.86 for MRI (8 studies), and 0.92 and 0.89 for US (4 studies), respectively. In studies incorporating clinical data alongside these primary imaging modalities, DP-based models showed modest improvements of 0.02–0.03 in median and mean AUC, MRI-based models demonstrated a more substantial increase of 0.04 in the median AUC, whereas US-based models exhibited minimal changes upon integrating clinical data. Additionally, acquiring imaging data at one more supplementary time point (early-, mid-, or post-NAST) beyond the baseline (pre-NAST) significantly increased median and mean AUC values from 0.82 to 0.91. However, incorporating more than one additional imaging acquisitions did not further enhance predictive performance. Key summary statistics of the selected 51 studies, including cohort sizes, pCR rates, and AUC distributions across all study categories can be found in supplementary data (Supplementary File 2).

MRI-based deep learning approaches for predicting pCR

Focusing specifically on the 20 studies identified in this review that utilized MRI as the primary data modality for DL-based prediction (Table 1), one study employed a retro‐prospective [27] design while all others were retrospective. Ten were multicenter studies, and nine among these leveraged the publicly available I-SPY1 [14, 44, 54–57] or I-SPY2 [23, 45, 46] cohorts as the primary study population or as additional external validation sets. The remaining ten studies were single-center, with one incorporating I-SPY2 [49] data for external validation. The majority of investigations encompassed nearly all BC subtypes, whereas one study exclusively focused on TNBC [27], and another focused on both TNBC and HER2+ BC [31]. Nearly all of the studies employed T1-based dynamic contrast-enhanced (DCE)-MRI. Eight studies employed only MRI for prediction, seven of them using end-to-end DL. Three of these studies [42, 55, 58] utilized only pre-NAST MRI, achieving an AUC range of 0.72–0.85 for pCR prediction, while the remaining five studies [27, 49, 51, 59, 60] incorporated MRI scans from the early-to-mid treatment phase, reporting an improved AUC range of 0.81–0.97. Eleven studies incorporated clinical variables alongside MRI data, utilizing demographic factors (age), tumor biological characteristics (ER, PR, HER2 status, Ki67), tumor staging (TNM), histopathological grading, menopausal status, and type of NAST regimen as predictive features, achieving AUC values ranging from 0.70 to 0.93. Across these MRI‐based investigations three main patterns of AUC improvement were confirmed also by intra-study comparisons. First, transitioning from radiomics to DL tends to improve performance, although head-to-head comparisons were limited [41, 61]. Second, combining MRI features with clinical data yields a performance gain over unimodal imaging approaches [14, 19, 31, 41, 44–46, 54, 56, 57, 61]. Finally, leveraging multiple time points (pre‐, early‐, mid- or post‐treatment) further raises predictive accuracy [14, 41, 45, 46, 48, 56].

Table 1.

Summary of studies using breast magnetic resonance imaging (MRI) as the primary input for pathologic complete response (pCR) prediction models

Author, year	Citation Nr	Study design	BC subtype	Data modalities	Longitudinal imaging	AI techniques	Nr of Pts	Held-out cohort	Overall pCR Rate (%)	Best validation AUC
Kim et al., 2024	31	Retrospective Single center	HER2+ BC and TNBC	MRI and Clinical	No	CNN end-to-end	852	Yes	36.4	0.70
Liu et al., 2020	55	Retrospective Multicenter (P)	Mixed	MRI only	No	CNN end-to-end	131	No	30.5	0.72
Massafra et al., 2022	57	Retrospective Multicenter (P)	Mixed	MRI and Clinical	No	CNN followed by CML	225	Yes	29.3	0.80
Comes et al., 2024	49	Retrospective Single center (P)	Mixed	MRI only	Yes	TRA end-to-end	106	Yes	36.8	0.81
Comes et al., 2024	58	Retrospective Single center	Mixed	MRI only	No	CNN end-to-end	120	No	34.2	0.82
Dammu et al., 2023	14	Retrospective Multicenter (P)	Mixed	MRI and Clinical	Yes	CNN end-to-end	155	No	27.1	0.83
Jing et al., 2024	45	Retrospective Multicenter (P)	Mixed	MRI and Clinical	Yes	CNN end-to-end	624	No	34.1	0.83
Peng et al., 2022	61	Retrospective Single center	Mixed	MRI and Clinical	No	CNN end-to-end	356	No	23.3	0.83
Zhou et al., 2023	27	Retro-prospective Single center	TNBC	MRI only	Yes	CNN end-to-end	210	Yes	48.1	0.83
Hao et al., 2023	42	Retrospective Single center	Mixed	MRI only	No	CNN end-to-end	442	Yes	21.5	0.85
Gao et al., 2024	23	Retrospective Multicenter (P)	Mixed	MRI, MG and Clinical	Yes	CNN end-to-end	3352	Yes	23.5	0.88
Verma et al., 2023	54	Retrospective Multicenter (P)	Mixed	MRI and Clinical	Yes	CNN end-to-end	121	Yes	30.6	0.88
Joo et al., 2021	19	Retrospective Single center	Mixed	MRI and Clinical	No	CNN end-to-end	536	Yes	24.8	0.89
Comes et al., 2021	56	Retrospective Multicenter (P)	Mixed	MRI and Clinical	Yes	CNN followed by CML	134	Yes	27.6	0.90
Wang et al., 2025	44	Retrospective Multicenter (P)	Mixed	MRI and Clinical	No	CNN end-to-end	281	Yes	41.3	0.90
El Adoui et al., 2020	59	Retrospective Single center	Mixed	MRI only	Yes	CNN end-to-end	56	Yes	35.7	0.91
Huang et al., 2023	51	Retrospective Multicenter	Mixed	MRI only	Yes	CNN followed by CML	1262	Yes	34.9	0.93
Li et al., 2023	41	Retrospective Single center	Mixed	MRI and Clinical	Yes	CNN followed by CML	95	Yes	25.3	0.93
Liu et al., 2025	46	Retrospective Multicenter (P)	Mixed	MRI and Clinical	Yes	CNN end-to-end	385	Yes	33.8	0.93
Qu et al., 2020	60	Retrospective Single center	Mixed	MRI only	Yes	CNN end-to-end	302	Yes	43.7	0.97

BC, Breast cancer; HER2+ BC, HER2-positive breast cancer; TNBC, Triple negative breast cancer; MG, Mammography; Clinical, Clinical data; AI, Artificial intelligence; CNN, Convolutional neural network; TRA, Transformer architecture; CML, Classical machine learning; Nr of Pts, Number of patients; (P), a Public dataset was involved; AUC, Area Under the Receiver Operating Characteristic Curve

With regards to neural network architectures, several studies employed custom-designed CNN, typically using end-to-end DL strategies for prediction, achieving AUC values ranging from 0.72 to 0.97 [14, 27, 44, 54, 55, 58–60]. Other investigations adopted pre-designed CNN architectures such as AlexNet [56, 57], ResNet [19, 23, 31, 42, 45, 51, 61], VGG [46], or hybrid models [41], reporting AUC values between 0.70 and 0.93. Among these, some studies trained these architectures de novo on their datasets, while others utilized pretrained CNNs leveraging transfer learning from large external datasets. Although pretrained CNN models were predominantly applied within end-to-end DL frameworks, some studies used CML methods to ensemble or integrate features extracted from one or more data modalities. Additionally, integration of transformer-based attention mechanisms or LSTM modules into CNN architectures was explored in several studies, and notably, one study exclusively implemented a Vision Transformer architecture [49].

Predictive value of deep learning models using ultrasound data

Twelve studies employed US as the primary data modality for DL-based prediction of pCR (Table 2). Among these, six studies were single-center, while eight utilized multicentric datasets. Most studies (n = 10) were retrospective, with two single-center studies designed prospectively [43, 62]. Investigated cohorts predominantly comprised mixed BC subtypes, except for one study specifically targeting HER2+ BC [29]. All included studies utilized two-dimensional (2D) US, primarily grayscale B-mode. Some investigations incorporated multiparametric US approaches, including strain elastography and various plane acquisitions (longitudinal, transverse, and largest cross-sectional planes). Four studies utilized exclusively US data for pCR prediction [29, 43, 47, 63], employing imaging acquired at pre-, early-, and/or mid-NAST, reporting AUC values ranging from 0.76 to 0.96. Six studies integrated clinical data with US imaging to form bimodal prediction models. Among these, half were limited to pre-NAST imaging acquisitions [52, 62, 64] and half integrated post-NAST imaging [48, 50, 53], with respective AUC ranges of 0.83–0.91 and 0.89–0.84. Two additional studies adopted a trimodal design, integrating US, DP, and clinical data, achieving AUC of 0.79 when only pre-NAST US was used [39] and AUC of 0.88 when early-NAST US was integrated [20]. Notably, eight studies out of 12 utilized longitudinal US assessments, combining pre-NAST imaging with subsequent early-, mid-, or post-NAST acquisitions.

Table 2.

Summary of studies using breast ultrasound (US) imaging as the primary input for pathologic complete response (pCR) prediction models

Author, year	Citation nr	Study design	BC subtype	Data modalities	Longitudinal imaging	AI techniques	Nr of Pts	Held-out cohort	Overall pCR Rate (%)	Best validation AUC
Gu et al., 2023	48	Retrospective Multicenter	Mixed	US and Clinical	Yes	TRA end-to-end	484	Yes	34.3	0.90
Gu et al., 2024	62	Prospective Single center	Mixed	US and Clinical	No	CNN end-to-end	170	Yes	37.1	0.88
Guo et al., 2024	20	Retrospective Single center	Mixed	US, DP and Clinical	Yes	TRA end-to-end	596	Yes	26.0	0.88
Jiang et al., 2021	50	Retrospective Multicenter	Mixed	US and Clinical	Yes	CNN followed by CML	592	Yes	31.8	0.94
Jiang et al., 2024	39	Retrospective Multicenter	Mixed	US, DP and Clinical	No	CNN followed by CML	311	Yes	16.7	0.79
Liu et al., 2022	29	Retrospective Multicenter	HER2+ BC	US only	Yes	CNN end-to-end	393	Yes	29.0	0.96
Liu et al., 2025	53	Retrospective Single center	Mixed	US and Clinical	Yes	CNN followed by CML	243	Yes	28.4	0.89
Tong et al., 2023	47	Retrospective Multicenter	Mixed	US only	Yes	TRA end-to-end	484	Yes	34.3	0.90
Wang et al., 2024	52	Retrospective Single center	Mixed	US and Clinical	No	CNN followed by CML	155	Yes	25.2	0.91
Xie et al., 2022	63	Retrospective Single center	Mixed	US only	Yes	CNN end-to-end	114	No	34.2	0.94
You et al., 2024	64	Retrospective Multicenter	Mixed	US and Clinical	No	CNN end-to-end	1409	Yes	37.6	0.83
Zhang et al., 2024	43	Prospective Single center	Mixed	US only	Yes	CNN end-to-end	57	Yes	50.9	0.76

BC, Breast cancer; HER2+ BC, HER2-positive Breast Cancer; DP, Digital Pathology; Clinical, Clinical data; AI, Artificial intelligence; CNN, Convolutional neural network; TRA, Transformer architecture; CML, Classical machine learning; Nr of Pts, Number of patients; (P), a Public dataset was involved; AUC, Area Under the Receiver Operating Characteristic Curve

Regarding feature extraction methodologies, all 12 studies leveraged DL techniques to various extent; however, three also incorporated quantitative radiomic parameters [50, 52, 53]. DL architectures varied across studies: six employed customized de novo CNNs [29, 43, 63] or transformer-based models [20, 47, 48] in end-to-end DL pipelines, consistently utilizing pre-NAST imaging alongside at least one subsequent US acquisition, with reported AUCs ranging from 0.76 to 96. Five studies utilized pre-trained models (DenseNet [50, 62], ResNet [39], DenseNet-ResNet combination [53], VGG [52]), integrating features via classical ML methods for final predictions; of these, three studies relied exclusively on pre-NAST imaging, reporting AUCs from 0.79 to 81, while three studies included longitudinal imaging with AUC ranging from 0.89 to 0.84. One study implemented a hybrid approach, combining pre-trained and custom-built DL model components, yielding an AUC of 0.83 [64].

Studies based on other radiological and functional imaging

Five studies utilizing imaging modalities beyond MR and US were identified, all including mixed BC subtypes (Table 3). Two single-center studies from the same research group specifically utilized MG alone for predicting pathological complete response (pCR). The first study, with a retro-prospective design, applied a pretrained transformer for tumor detection and a ResNet model for feature extraction and classification on pre-NAST 2D digital mammograms, achieving an AUC of 0.71 [37]. The subsequent prospective study leveraged digital breast tomosynthesis (3D) images acquired at three distinct time points (pre-NAST, after two NAST cycles, and post-NAST) [36]. Using a pretrained 3D ResNet coupled with an attention module, longitudinal imaging significantly enhanced predictive performance, resulting in a final AUC of 0.83.

Table 3.

Summary of studies using computed tomography (CT), mammography (MG) or positron emission tomography (PET) imaging as the primary input for pathologic complete response (pCR) prediction models

Author, year	Citation nr	Study design	BC subtype	Data modalities	Longitudinal imaging	AI techniques	Nr of Pts	Held-out cohort	Overall pCR Rate (%)	Best validation AUC
Bulut et al., 2023	22	Retrospective Single center	Mixed	PET only	No	CNN end-to-end	31	No	29.0	0.90
Förnvik et al., 2024	36	Prospective Single center	Mixed	MG only	Yes	CNN end-to-end	149	Yes	22.1	0.83​
Rezaeijo et al., 2021	34	Retrospective Single center (P)	Mixed	CT only	Yes	CNN end-to-end	121	Yes	47.9	1.00
Skarping et al., 2022	37	Retro-prospective Single center	Mixed	MG only	No	CNN end-to-end	453	Yes	21.0	0.71
Tan et al., 2022	35	Retrospective Multicenter	Mixed	CT and Clinical	No	CNN end-to-end	324	Yes	22.8	0.77

BC, Breast cancer; Clinical, Clinical data; AI, Artificial intelligence; CNN, Convolutional neural network; Nr of Pts, Number of patients; (P), a Public dataset was involved; AUC, Area Under the Receiver Operating Characteristic Curve

Two additional retrospective studies utilized breast CT within end-to-end DL frameworks for pCR prediction. The first employed publicly available imaging data (QIN-Breast dataset from The Cancer Imaging Archive), with CT images obtained pre-NAST and after the first treatment cycle [34]. An ensemble deep transfer learning approach integrating pretrained ResNet and DenseNet architectures demonstrated exceptionally high predictive accuracy, achieving an AUC of 1.00 in internal testing [34]. The second retrospective, multicentric study combined pre-NAST CTI and clinical data [35]. Employing both quantitative radiomics and DL-based feature extraction through a custom MultiResUnet3D architecture, the study showed that combining imaging features with clinical data via logistic regression improved predictive performance to an AUC of 0.77, superior to either modality alone.

Lastly, a smaller retrospective single-center study assessed the predictive utility of pre-NAST breast 18F-fluorodeoxyglucose positron emission tomography/computed tomography (PET/CTI) [22]. Employing a pretrained ResNet architecture in an end-to-end DL framework, this investigation reported an AUC of 0.90; however, evaluation was limited to cross-validation.

Leveraging digital pathology imaging through deep learning

Thirteen studies utilized pre-NAST DP imaging to develop DL models for predicting pCR (Table 4). All studies were retrospective, with diverse BC subtype representations: seven examined mixed subtypes, four focused specifically on TNBC [24–26, 28], one on HER2+ BC [30], and another on HER2- BC (including both HR+/HER2− BC and TNBC cases) [33]. Eight studies used DP as the only data input [24–26, 33, 40, 65–67], reporting area AUC values ranging from 0.71 to 0.96 (median 0.80). Four studies combined DP-derived features with clinical data [28, 30, 68, 69], with reported AUC ranges between 0.75 and 0.89 (median 0.81). A single study integrated DP with CT, achieving an AUC of 0.93.

Table 4.

Summary of studies using digital pathology (DP) imaging as the primary input for pathologic complete response (pCR) prediction models

Author	Citation nr	Study design	BC subtype	Data modalities	Longitudinal imaging	AI techniques	Nr of Pts	Held-out Cohort	Overall pCR Rate (%)	Best validation AUC
Bhattarai et al., 2024	25	Retrospective Single center	TNBC	DP only	No	CNN followed by CML	76	No	57.9	0.71
Fanucci et al., 2023	33	Retrospective Multicenter	HER2- BC	DP only	No	CNN followed by CML	113	No	29.2	0.71
Aswolinskiy et al., 2023	67	Retrospective Multicenter (P)	Mixed	DP only	No	CNN followed by CML	973	Yes	24.0	0.72
Krishnamurthy et al., 2023	28	Retrospective Single center	TNBC	DP and Clinical	No	CNN followed by CML	243	Yes	37.4	0.75
Li et al., 2022	66	Retrospective Multicenter	Mixed	DP only	No	CNN followed by CML	874	Yes	25.6	0.78
Zeng et al., 2024	69	Retrospective Multicenter	Mixed	DP and Clinical	No	CNN followed by CML	440	Yes	28.2	0.78
Li et al., 2022	65	Retrospective Multicenter	Mixed	DP only	No	CNN followed by CML	1035	Yes	24.5	0.79
Li et al., 2024	26	Retrospective Single center	TNBC	DP only	No	CNN end-to-end	105	No	45.7	0.82
Wang et al., 2025	30	Retrospective Single center	HER2+ BC	DP and Clinical	No	TRA followed by CML	399	No	59.1	0.84
Li et al., 2021	68	Retrospective Single center	Mixed	DP and Clinical	No	CNN followed by CML	540	Yes	18.9	0.89
Saednia et al., 2023	40	Retrospective Single center	Mixed	DP only	No	CNN end-to-end	207	Yes	25.1	0.89
Zhang et al., 2023	38	Retrospective Single center	Mixed	DP and CT	No	CNN followed by CML	211	Yes	38.9	0.93
Duanmu et al., 2022	24	Retrospective Single center	TNBC	DP only	No	CNN end-to-end	73	No	58.9	0.96

BC, Breast cancer; HER2+ BC, HER2-positive Breast cancer; HER2- BC, HER2-negative Breast Cancer; TNBC, Triple negative breast cancer; CT, Computed tomography; Clinical, Clinical data; AI, Artificial intelligence; CNN, Convolutional neural network; TRA, Transformer architecture; CML, Classical machine learning; Nr of Pts, Number of patients; (P), a Public dataset was involved; AUC, Area Under the Receiver Operating Characteristic Curve

Twelve investigations used WSIs, primarily captured with specialized scanners: eight at 40× magnification, three at 20× magnification, and one at 10× magnification. Ten of these studies utilized H&E staining exclusively, while two incorporated additional stains with proliferation markers (Ki67 and PH3) [24, 25]. One study deviated from WSIs, capturing selected fields via microscopy (20× objective) of HER2-stained slides, and employed a vision transformer for feature extraction followed by logistic regression, reporting an AUC of 0.84 through cross-validation without independent testing. The remaining 12 studies relied on CNN-based feature extraction. Three studies employed ResNet architectures, either pretrained followed by CML [38, 66] or trained de novo in end-to-end DL pipeline [24]. Two studies utilized Inception architectures combined with CML classifiers [65, 68], while two other investigations leveraged VGG architectures (one pretrained, followed by CML prediction [26, 69], the other trained de novo with DL prediction [26]). Other employed architectures included pretrained U-Net [67] and de novo trained Mask R-CNN [25] coupled with CML, and de novo trained CoatNet using an end-to-end DL approach [40]. Two studies introduced novel customized CNN architectures paired with CML classifiers [28, 33]. In summary, six studies employing pre-designed CNN architectures trained de novo reported AUC values ranging from 0.71 to 0.96 (median 0.86). Four studies utilizing pretrained CNN models achieved AUCs from 0.72 to 0.93 (median 0.78), whereas two studies employing custom-designed models achieved AUCs of 0.71 and 0.75.

Deep learning applied to molecular omics and clinical text data

We identified only one study falling in the category of investigations that applied DL on tumor molecular data (i.e. genomics, epigenomics, transcriptomics and/or proteomics) (Table 5). In particular, this study involved 368 patients diagnosed with HER2-low BC from the I-SPY2 Trial [32]. Investigators applied DL on transcriptomic data (mRNA) as the primary data modality, and integrated it with clinical data (including textual elements from pathology reports) obtaining a final classifier with an excellent AUC of 0.92 in predicting pCR. The DL architecture consisted in a customized fully connected network with 6 hidden layers, that fused differential gene expression patterns with clinical elements in an end-to-end DL fashion.

Table 5.

Summary of a study using RNA sequencing data as the primary input for pathologic complete response (pCR) prediction models

Author	Citation nr	Study design	BC subtype	Data modalities	Longitudinal imaging	AI techniques	Nr of Pts	Held-out cohort	Overall pCR Rate (%)	Best validation AUC
Li et al., 2023	32	Retrospective Multicenter (P)	HER2-low BC	RNA and Clinical	No	FCN end-to-end	368	Yes	26.6	0.92

BC, Breast cancer; Clinical, Clinical data; AI, Artificial intelligence; FCN, Fully connected neural network; Nr of Pts, Number of patients; (P), a Public dataset was involved; AUC, Area Under the Receiver Operating Characteristic Curve

In this systematic review we did not identify any investigation that employed pre-surgical clinical data as the primary data input modality in DL-based predictors of pCR in BC patients undergoing NAST.

Discussion

This systematic review synthesizes the landscape of DL applications across multiple data types to predict pCR to NAST in BC. Our analysis of 51 studies reveals a rapidly evolving field with considerable potential, particularly for multimodal approaches, while also highlighting significant methodological heterogeneity and areas for future development.

We confirm the growing interest in deep learning for pCR prediction, primarily using radiological data (MRI and US), followed by DP imaging (Fig. 2). Multimodal models generally performed better than single-modality models in the same study; however, formal statistical comparison was not reported. A compelling demonstration of how true multimodality and end-to-end DL can work in concert was published recently. Using a prospective, multicentre cohort of 1004 patients, the DL model fused pretreatment MRI, WSI pathology images and structured clinical variables inside one unified CNN-based pipeline, and achieved an external-test AUC of 0.88 and a prospective-test AUC of 0.91, significantly higher than any of its single-modality counterparts [70]. In our systematic review, clinical data integration was the most common strategy yielding some performance improvements, especially alongside MRI. Among the 51 studies, purely end-to-end CNN pipelines were not only the most common (n = 21) but also the most accurate (median AUC 0.88). When the CNN served only as a feature extractor and the final classifier was performed using conventional ML, performance dipped appreciably (median 0.80). Adding attention blocks to a CNN backbone yielded an intermediate benefit (median 0.85), while the small pool of transformer-only models posted a slightly higher median (0.89). A similar pattern emerged across input modalities: MRI-only models reached a median AUC of 0.84, which climbed to 0.88 once routine clinical variables were fused; DP rose modestly from 0.79 to 0.81 with the same addition, whereas US maintained high accuracy regardless of fusion (0.92 vs 0.90), probably because of the frequent addition of mid-treatment inputs. For instance, one of the studies we included in the systematic review extracted features from biopsy WSIs using a VGG16 CNN-architecture and, after fusing them with routine clinicopathological variables in an SVM, improved AUC from 0.79 to 0.84 in the training set and from 0.71 to 0.78 on an external cohort [69]. Beyond the modalities analysed in this review, additional imaging inputs are possible. An example is functional optical imaging which is also being explored in this context: a study fused co-registered US imaging with diffuse optical tomography (DOT) in a dual-input transformer, plus clinical variables, and achieved an AUC of 0.96 (vs 0.87 for US and 0.82 for DOT alone), underscoring that additional, complementary imaging streams can further enhance multimodal DL performance [71]. Taken together, these trends reinforce two messages already apparent in Fig. 3 and Supplementary Fig. 1: (i) end-to-end deep learning usually outperforms hybrid pipelines, and (ii) incremental clinical context can improve performance, depending on the imaging modality involved. Incorporating longitudinal imaging during NAST boosted accuracy over baseline imaging alone, underscoring the value of capturing tumor dynamics during treatment. However, more than two imaging time points did not offer further gains. DL models achieved a median AUC of 0.88, with 35.3% reporting ≥ 0.90, indicating strong discrimination for pCR. Because many AUCs were reported from retrospective single-centre cohorts, these values should be interpreted cautiously. Median and mean AUC values were slightly higher (by 0.01–0.04) in studies with intermediate cohort sizes (156–400 patients) and those with intermediate-to-high pCR rates (above 27%), highlighting a limitation: model performance appears to degrade when pCR rates fall below 25%(Supplementary File 2). Interestingly, this trend toward higher AUC was even more pronounced (by 0.04–0.05) among studies that validated results independently on a held-out test cohort (Supplementary Fig. 2). Nonetheless, within these studies, the AUC values derived from independent testing cohorts (which we report in this study) were consistently lower compared to cross-validation AUC obtained from the training cohorts. From an assisted clinical decision-making perspective, an externally validated model that reaches an AUC of about 0.90 means that in roughly nine out of ten patient pairs the algorithm will correctly rank the likelihood of achieving pCR, giving the breast oncologist a reliable probability estimate on which to tailor (or even forego) neoadjuvant therapy intensity.

Fig. 3.

Bubble plot displaying the median Area Under the Receiver Operating Characteristic Curve (AUC) of studies grouped based on the deep learning (DL) architecture utilized. Each circle represents a set of studies, with relative dimension representing the number of studies included. CNN, Convolutional neural network; CNN + , Convolutional neural network augmented with additional modules like Attention mechanisms, etc.; TRA, Transformer architecture; CML, Classical machine learning

CNNs are the cornerstone, used in nearly 90% of studies for feature extraction within end-to-end pipelines. Many studies augment CNNs with attention, transformers, or LSTMs for longitudinal data, and a small but growing number rely solely on transformer architectures. Although end-to-end deep learning predominated, a minority (18 studies) used hybrid models with DL for feature extraction and conventional ML for data integration and final prediction. The optimal strategy remains unclear and likely depends on data characteristics and availability. Our findings align with broader trends of predictive models in oncology: moving from traditional statistical models or ML with handcrafted features (like radiomics) to DL enhances discriminative power, especially when integrating multiple data types. A prior meta-analysis of MRI-based ML for NAST response reported higher AUC for DL (0.92) than classical ML + radiomics (0.85), and showed that adding clinical or histopathologic information improved performance (0.90 vs. 0.82 for MRI alone) [72]. These results independently confirm the value of clinical integration and the benefit of multimodal data. Compared to earlier reviews focused on unimodal radiomics or classical ML, this review highlights the shift toward sophisticated, data-driven feature extraction and integration, while underscoring persistent challenges in standardization and validation.

Despite promising AUCs, most studies (90.2%) were retrospective with potential selection bias, which often inflates performance because favourable cases are unintentionally over-represented. Moreover, often studies used the same public datasets, limiting causal inference, generalizability, and introducing risk of overfitting. Although 76.5% included independent testing on held-out cohorts, they were often from the same institution and true external validation remains scarce. Methodological heterogeneity—in patient cohorts, BC subtypes, NAST regimens, imaging protocols, DL architectures, fusion strategies, pCR definitions, and performance reporting—hinders direct comparisons and precludes meta-analysis. Notably, molecular multi‑omics data remain underutilized: we found only one transcriptomics-based DL study for pCR prediction [32]. Yet prior work has shown many signals highlighting the potential predictive power of molecular multi-omics in the DL framework. In fact, DL has been successfully applied directly on raw cancer DNA sequence data for BC type prediction and driver gene identification [73], on epigenomic data (DNA methylation) for classifying BC subtypes [74], and in sophisticated efforts to integrate multi-omics data (including copy number alterations, mutations, methylation, RNA, and protein expression)—sometimes combined with clinical data—to predict broader outcomes like survival and drug response in BC [75–77]. More specifically on breast cancer, several recent molecular biology work underscores just how much actionable signal is being left unexploited. Metabolic and signalling drivers such as the RRP9-JUN/AKT axis [78] and ZFP64-mediated glycolytic reprogramming [79] have been shown to accelerate tumour growth, while a circSpdyA-encoded 127-aa micro-peptide directly binds FASN to boost lipogenesis and dampen NK-cell cytotoxicity [80]. Immune-modulatory molecules like MOSPD1, whose silencing augments response to anti-PD-L1 therapy [81], and epigenetic regulators such as the PML1-WDR5 complex [82] or the hypoxia-activated XBP1s/HDAC2/EZH2 cassette that represses ΔNp63α [83] further illustrate the breadth of mechanistic biomarkers now charted. Even single-patient multi-omic atlases, e.g. the claudin-low luminal B breast cancer case carrying high TMB and APOBEC signatures [84], show how layered genomic, epigenomic and immune biomarkers data can contain predictive information to complement other data inputs. Incorporating such mechanistic features, whether as raw omics tensors or distilled pathway scores, into future multimodal DL frameworks could enhance biological interpretability, synergise with WSI‐derived morphometry, and ultimately refine personalised therapy selection. Similarly, the complete absence of studies using DL primarily on pre-treatment clinical text data (e.g., EMR data, pathology/radiology reports, clinical notes) represents a missed opportunity. While classical ML approaches applied to structured clinicopathological data have shown moderate success in predicting pCR, achieving AUCs around 0.71–0.79 in recent studies [85, 86], these methods typically rely on pre-selected features. The potential of leveraging the richer, unstructured information within clinical narratives using advanced DL techniques like LLMs remains largely underexplored for pre-treatment pCR prediction in this context. Although one study demonstrated the feasibility of using LLMs to extract pCR information from post-surgery pathology reports [87], and the general power of LLMs for deep phenotyping from clinical text is established [88], no eligible study in our review applied these powerful text-processing models to predict pCR using pre-NAST clinical narratives. Critically, no study attempted to integrate all four major data domains (radiological, pathological, molecular, clinical), which would be necessary to build truly comprehensive predictive models.

Current DL models show strong potential but are not yet ready for routine clinical use. Their retrospective design, lack of external validation, and methodological variability must be addressed to ensure reliability in clinical decisions. Future research should prioritize prospective, multi‑center trials for robust validation, standardize MRI protocols, pCR definitions, and feature‑extraction methods, and adopt unified outcome measures to improve reproducibility. True multimodal integration should incorporate underutilized modalities using sophisticated fusion techniques beyond simple concatenation. This holistic approach holds substantial promise for improved predictive accuracy and personalized patient management. Moreover, ongoing exploration and application of advanced DL architectures, including transformers, graph neural networks and foundation models [89–91], should be encouraged.

Equally important is the development and implementation of explainable AI methods, which were formally included in the design by few of the studies included in the review [23, 58]. A recent example of good practice is in fact a Multi-modal Response Prediction model, which used Integrated Gradients to rank modality-specific contributions and confirmed that histopathological subtype and MRI-derived features were the principal drivers of its predictions, thereby providing clinicians with transparent, clinically plausible explanations for each case [23]. Explainability is critical for moving beyond the "black box" nature of DL models, fostering clinical trust, and clarifying the specific factors influencing predictions. Robust model evaluation beyond AUC is also critical. Decision curve analysis, model calibration, and assessments of clinical impact on patient outcomes and cost-effectiveness should be incorporated to determine whether high AUCs translate into clinically useful decision thresholds. Finally, fostering multi-institutional collaboration and employing privacy-preserving methods such as federated learning could overcome limitations associated with single-center studies. Indeed, federated learning approaches have already demonstrated feasibility in this domain, having been used to build predictors of NAST response by applying machine learning to WSIs and clinical data across different institutions while maintaining data privacy behind their respective firewalls [92]. This collaborative approach would facilitate model training on larger, more diverse datasets, thus enhancing generalizability and potential clinical utility.

The strengths of this review include a systematic search strategy across multiple databases and a focused synthesis of current methodologies, performance, and gaps. While another systematic review published in 2022 also explored DL for pathological complete response prediction, its focus was limited to MRI and clinical data [93], whereas the present review encompasses a broader range of modalities, offering a more comprehensive and updated perspective. However, its main limitation is the inability to perform a quantitative meta-analysis due to the significant heterogeneity across the included studies, necessitating a narrative synthesis without a single pooled estimate of effect size.

Conclusion

This systematic review highlights the potential of DL, particularly when applied multimodally, for predicting pCR to NAST in BC. This review demonstrates that multimodal DL approaches can achieve high discriminative performance for pCR prediction in retrospective cohorts; however, their clinical utility remains unproven without prospective validation. Likewise, multimodal models hold the potential to redirect patients to more effective/tolerated (escalation/de-escalation) schedules, considering the growing pipeline of neoadjuvant treatments. The integration of multiple data sources, most commonly imaging and clinical data, along with the use of longitudinal assessments during therapy, often demonstrates improved predictive performance compared to unimodal or baseline-only approaches. Despite these promising results, the field is characterized by significant methodological heterogeneity, a predominance of retrospective studies, and a lack of rigorous external validation, which currently limits the clinical applicability of these models. Advancing the field towards clinical utility requires prioritizing prospective multi-center validation studies, fostering standardization in data acquisition and reporting, developing innovative methods for integrating underutilized data like multi-omics and clinical text, exploring “new generation” DL architectures, and embedding explainable AI techniques. Addressing these challenges is crucial for building robust, generalizable, and trustworthy models capable of facilitating personalized treatment strategies in BC care.

Author contributions

Eriseld Krasniqi: conceptualization, methodology, investigation, formal analysis, visualization, project administration, writing—original draft, writing—review and editing. Lorena Filomeno: methodology, investigation, writing—review and editing. Teresa Arcuri: methodology, investigation, writing—review and editing. Gianluigi Ferretti: writing—review and editing. Simona Gasparro: writing—review and editing. Alberto Fulvi: writing—review and editing. Arianna Roselli: writing—review and editing. Loretta D’Onofrio: writing—review and editing. Laura Pizzuti: methodology, writing—review and editing. Maddalena Barba: writing—review and editing. Marcello Maugeri-Saccà: supervision, writing-review and editing. Claudio Botti: writing—review and editing. Franco Graziano: writing—review and editing. Ilaria Puccica: writing—review and editing. Sonia Cappelli: writing-review and editing. Fabio Pelle: writing-review and editing. Flavia Cavicchi: writing—review and editing. Amedeo Villanucci: writing—review and editing. Ida Paris: writing-review and editing. Fabio Calabrò: writing-review and editing. Sandra Rea: writing-review and editing. Maurizio Costantini: writing—review and editing. Letizia Perracchio: writing—review and editing. Giuseppe Sanguineti: writing—review and editing. Silvia Takanen: writing—review and editing. Laura Marucci: writing—review and editing. Laura Greco: methodology, writing—review and editing. Rami Kayal: methodology, writing-review and editing. Luca Moscetti: writing-review and editing. Elisa Marchesini: writing—review and editing. Nicola Calonaci: methodology, investigation, formal analysis, visualization, writing—review and editing. Giovanni Blandino: conceptualization, supervision, writing—review and editing. Giulio Caravagna: methodology, investigation, formal analysis, visualization, supervision, writing-review and editing. Patrizia Vici: conceptualization, investigation, funding acquisition, supervision, writing—review and editing. All authors read and approved the final manuscript.

Funding

This work was financially supported through funding from the institutional “Ricerca Corrente” granted by the Italian Ministry of Health.

Availability of data and materials

All data generated or analysed during this study are included in this published article and its supplementary data files.

Declarations

Ethics approval and consent to participate.

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.