Ultrasound-based predictive models for response to neoadjuvant chemotherapy in breast cancer: a systematic review and meta-analysis Authors: Cuo Yi, Ting Xu, Rui You, Qiang Zhang, Ruihan Liu, Xi Yang Published: 2026 ABSTRACT AbstractObjectiveThis meta-analysis aimed to assess the diagnostic efficacy of ultrasound-based prediction models in evaluating the response to neoadjuvant chemotherapy (NAC) in breast cancer patients.MethodsA systematic review was conducted using PubMed, Embase, and Web of Science from inception to May 2025. We focused on studies evaluating ultrasound-based prediction models for NAC response in breast cancer. Two investigators independently performed study selection and data extraction according to predefined eligibility criteria. The risk of bias was assessed using the Prediction Model Risk of Bias Assessment Tool (PROBAST).Meta-analysis was performed using Stata 14.0 software, with the area under the curve (AUC) and its 95% confidence interval (CI) as the primary measure of diagnostic accuracy.ResultsFourteen retrospective studies with a total of 35 prediction models were included. PROBAST assessment indicated a high overall risk of bias in most studies (10 out of 14). The pooled AUC for all ultrasound-based prediction models in predicting NAC response was 0.84 (95% CI: 0.82, 0.86). However, substantial heterogeneity was observed among the studies (I2 = 69.7%, p < 0.001). Subgroup analyses suggested that model features and the prediction outcomes were major sources of heterogeneity. In the test cohort, the model combining ultrasound with clinical features demonstrated the best performance in predicting pathologic complete response (pCR), with a pooled AUC of 0.88 (95% CI: 0.84–0.92) and no heterogeneity (I2 = 0.0%). Sensitivity analyses confirmed the robustness of the findings.Although the funnel plot appeared symmetrical, both Begg's and Egger's tests indicated publication bias (p < 0.01). Notably, 10 of 14 included studies (71.4%) were at high overall risk of bias, and all studies were single-center and conducted in China, which substantially limits the immediate clinical translation of these findings.ConclusionUltrasound-based prediction models show encouraging discriminative ability for predicting pCR (pooled AUC 0.84, 95% CI 0.82–0.86), suggesting preliminary potential.However, given the high risk of bias, substantial heterogeneity, and geographic limitation (all studies from China), the current evidence should be considered exploratory and hypothesis-generating rather than practice-changing. Future prospective, multicenter, and rigorously designed studies are needed to confirm these findings.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12885-026-16330-6.Keywords: Breast Neoplasms, Neoadjuvant Chemotherapy, Ultrasonography, Prediction Model, Systematic Review, Meta-analysis Introduction Breast cancer is the most commonly diagnosed cancer in women worldwide, with approximately 2.3 million new cases are diagnosed annually, accounting for 11.7% of all new cancer cases in 2020 [1]. This poses a significant threat to women's physical and mental health. Advances in early screening, diagnosis, and treatment have raised the five-year survival rate to exceeding 90% in high-income countries [2]. The overarching goal of contemporary oncology has evolved from merely extending survival to optimizing the quality of life. In this context, breast-conserving surgery (BCS) is a critical indicator of patient-centered care. However, the breast-conserving surgery rate in China remains low at about 21.9% [3], indicating that patients' quality of life needs to be significantly improved. NAC is the standard treatment for locally advanced breast cancer, aimed at reducing tumor stage, increasing the rate of breast-conserving surgery, and improving patient prognosis [4–6]. However, not all breast cancer patients benefit equally from NAC. Studies report that approximately 22.3% of patients achieve pCR after NAC [7]. About 47.4% achieve nodal pCR [8], meaning complete eradication of cancer cells in the lymph nodes—a state defined as the absence of lymph node metastasis (LNM). This wide variation highlights significant individual variation in treatment efficacy. The response to NAC critically influences subsequent clinical strategies, including the extent of surgery and adjuvant systemic therapy (e.g., chemotherapy, endocrine therapy) [9–11]. In addition, accurate prediction of NAC response could also inform locoregional radiotherapy decisions (e.g., nodal irradiation volume or de- escalation/escalation strategies), although this remains an investigational area. Therefore, developing accurate and non-invasive predictive models for NAC response before or during early treatment stages holds significant clinical value. Ultrasound is a fundamental imaging tool for breast cancer, widely used to evaluate patients before and after NAC because it is cost-effective, reproducible, and free of radiation [12–14]. However, its accuracy depends heavily on the operator and subjective interpretation of tumor features, which limits its precision in predicting pathological response [15, 16]. Studies report that the false-negative rate (FNR) of ultrasound for identifying pCR can be as high as 39.2% [17], indicating a need for improved sensitivity and specificity. Consequently, there is a clear need for a stable and objective clinical prediction model to reliably forecast NAC response in breast cancer. Several studies have explored predictive models that combine clinicopathological features with ultrasound imaging to quantify individual NAC responses [18–31]. However, current research has certain limitations. Firstly, the models and predictive indicators used vary significantly across studies, leading to considerable heterogeneity. Secondly, the overall diagnostic performance and optimal clinical application scenarios remain unclear. Moreover, systematic reviews and quantitative analyses that synthesize existing evidence are lacking. Unlike previous meta-analyses [32–35] that broadly evaluated conventional ultrasound modalities (grayscale, Doppler, CEUS, elastography) or focused on machine learning with earlier search cutoffs, our study provides an updated synthesis of ultrasound-based radiomics and machine learning prediction models (search up to May 2025). We also directly compare ultrasound-only versus combined-feature models, provide a rigorous PROBAST-based risk-of-bias assessment, explicitly discuss geographic limitations, and offer a tempered, hypothesis-generating conclusion. To address these gaps, this study conducts a meta-analysis of ultrasound-based prediction models for evaluating NAC response in breast cancer. Our objectives include determining the overall diagnostic performance of these models. We also aim to identify sources of heterogeneity through subgroup analyses and assess study quality, risk of bias, and result stability. The findings are intended to provide robust evidence for clinical practice and guide the development of superior predictive models. Methods Search strategy We searched the literature on PubMed, Embase, and Web of Science from their inception until May 2025 for relevant studies on ultrasound-based prediction models evaluating the response of breast cancer to NAC. The search strategy combined controlled vocabulary (e.g., MeSH terms) and free-text keywords. Specifically, the search terms including "Ultrasonography," "Deep learning," "Breast Neoplasms," "Drug Therapy," and their synonyms, were connected by Boolean operators such as AND and OR. The complete search strategies for all databases are provided in Supplementary File 1. Inclusion and exclusion criteria According to the PICOTS (Population, Intervention, Comparator, Outcome, Timing, Setting) framework, the inclusion and exclusion criteria for this study were as follows: P (Population): Breast cancer patients treated with NAC and subsequent surgery for pathological response assessment. P (Population): Breast cancer patients treated with NAC and subsequent surgery for pathological response assessment. I (Prediction Model): Ultrasound-based prediction models for the response to neoadjuvant chemotherapy. I (Prediction Model): Ultrasound-based prediction models for the response to neoadjuvant chemotherapy. C (Comparator): Histopathological evaluation of the surgical specimen as the reference standard for determining pCR, CR and/or axillary lymph node response. C (Comparator): Histopathological evaluation of the surgical specimen as the reference standard for determining pCR, CR and/or axillary lymph node response. (Outcome): AUC with its 95% CI as the metric for evaluating model discrimination. (Outcome): AUC with its 95% CI as the metric for evaluating model discrimination. T (Timing): Ultrasound examinations performed at pre-NAC and/or during or post-NAC. T (Timing): Ultrasound examinations performed at pre-NAC and/or during or post-NAC. S (Setting): Observational studies that develop, validate, or apply an ultrasound-based prediction model for NAC response. S (Setting): Observational studies that develop, validate, or apply an ultrasound-based prediction model for NAC response. The inclusion and exclusion criteria of this study were defined based on the PICOTS framework. Studies that used ultrasound-based prediction models to evaluate response to NAC in breast cancer were included. Model performance was assessed using the AUC along with its 95% CI. The exclusion criteria included: (1) publications in non-English languages; (2) studies that used only a training cohort without independent validation or testing; (3) publication types such as reviews, case reports, conference papers, meta-analyses, and letters; and (4) studies with incomplete or inaccessible data. Study selection and data extraction The references retrieved from various databases were imported into EndNote software to remove duplicate records. After that, two researchers independently screened the titles and abstracts to exclude studies that clearly did not meet the inclusion criteria. This process completed the initial screening. The full texts of the shortlisted studies were then thoroughly reviewed by the same two researchers. They independently determined the final list of included studies and cross-checked their results. Data extracted from the included studies included the authors, publication year, journal, study population country, study period, design, sample size, pathological subtypes of breast cancer, tumor stage, chemotherapy response criteria, and ultrasound techniques used. The full dataset extracted from the included studies is available in Supplementary File 4 (Table S1). Risk of bias assessment The risk of bias for each study was assessed using the PROBAST, which includes four domains and twenty signaling questions. Each domain was rated as high risk, low risk, or unclear. A model was rated as having a low overall risk of bias if all four domains were rated as low risk. Data analysis Statistical analysis was performed using Stata 14.0 software. The AUC along with its 95% CI was selected as the statistical metric and was used to evaluate the predictive ability of the models. The AUC from the test cohort was used preferentially for analysis. When the AUC from a test cohort was not available, that from the validation cohort was used instead. The AUC derived solely from training cohorts was excluded from the analysis to avoid overestimating model performance. Heterogeneity was assessed using the I2 statistics and Cochran's Q test. An I2 value < 50% or a Q test p-value > 0. 10 indicated no significant heterogeneity, in which case a fixed-effects model was applied. Conversely, an I2 value ≥ 50% or a p-value ≤ 0.10 indicated significant heterogeneity, leading to the use of a random-effects model. Discrimination performance was primarily assessed using the area under the receiver operating characteristic curve (AUC), a threshold-independent metric that summarizes a model’s ability to distinguish responders from non-responders across all possible classification thresholds. In the context of pCR prediction for NAC response, an AUC of 0.70–0.79 is generally considered acceptable, 0.80–0.89 as good, and ≥ 0.90 as excellent [36]. However, AUC alone is insufficient for clinical decision-making because it does not account for prevalence or the relative consequences of false-positive versus false-negative predictions. The clinical actionability of a model depends on the specific scenario: in de-escalation contexts (e.g., avoiding axillary dissection), high sensitivity (≥ 90%) is required to minimize missed true responders; in escalation contexts (e.g., recommending post-mastectomy radiation), high specificity is prioritized to avoid overtreatment. The pooled AUC of 0.84 (95% CI 0.82–0.86) falls within the ‘good’ range but does not, by itself, establish clinical utility without prospective validation of the model’s operating characteristics in real-world settings. Subgroup analysis was conducted to further explore the sources of heterogeneity among the studies. In addition to the prespecified subgroup analyses (data type, model features, prediction outcomes), we post hoc performed a subgroup analysis based on the ultrasound technique used (B-mode alone vs. B-mode + Doppler) to address reviewer feedback. Sensitivity analysis was performed to examine the robustness of the findings. Publication bias was evaluated using a funnel plot, Begg's test, and Egger's test. Results Literature screening results We conducted a literature search in PubMed, Embase, and Web of Science. We applied filters for English language and human subjects and identified 671 articles. After removing the 208 duplicate records, 423 articles were excluded based on title and abstract screening for not meeting the eligibility criteria. The remaining 40 articles were assessed through full-text review. Among them, 24 were excluded due to outcomes other than AUC, and 2 were excluded for not reporting the 95% CI of the AUC. Finally, 14 studies [18–31] were included in the meta-analysis. Figure 1 illustrates the study search and selection process. Flowchart of searching and screening steps for the studies Basic characteristics of the included literature The characteristics from the included studies are summarized in Supplementary File 4 (Table S1). All 14 studies were conducted in China and were retrospective in design. Sample sizes varied from 155 to 1220, totaling 5768 patients: 3549 in the training cohort, 1552 in the validation cohort, and 667 in the test cohort. Regarding model development pipelines, three studies implemented complete processes involving training, validation, and test phases. Nine studies did not include a test phase, while two proceeded directly to test without a validation step. The studies covered early, locally advanced, and advanced stages of breast cancer. Ten studies used molecular subtypes of breast cancer, three used histopathological breast cancer subtypes, and one study did not specify the pathological subtype. All studies focused on NAC. pCR was used as the treatment response indicator in nine studies, LNM in two studies, both pCR and LNM in another two studies, and one study used partial response (PR) according to the Response Evaluation Criteria in Solid Tumors (RECIST) criteria. In terms of data sources, nine studies utilized B-mode ultrasound only for model development, while five studies combined B-mode and Doppler ultrasound. Risk of bias assessment We assessed the risk of bias in 14 studies on prediction models for breast cancer chemotherapy response using the PROBAST tool. The results are summarized in Fig. 2. Bias risk assessment for breast cancer prediction models using PROBAST Tool The evaluation covered four domains: participants (D1), predictors (D2), outcome (D3), and analysis (D4), and an overall rating was assigned to each study. The results show that the outcome domain (D3) was well controlled, with the vast majority of studies (13 out of 14) rated as low risk in this aspect. The risk of bias in the predictors domain (D2) was moderate, with seven studies judged as low risk and only one as high risk. However, the participants (D1) and analysis (D4) domains emerged as methodological weaknesses. Ten studies were rated as high risk in the participants (D1) domain and six studies in the analysis (D4) domain, respectively, leading to an overall high risk of bias in most studies (10 out of 14). These findings suggest that future studies should focus on optimizing patient selection criteria and statistical analysis methods. Furthermore, the methodological heterogeneity identified here should be carefully considered when interpreting the present meta-analysis. Meta‑analysis The meta-analysis included 35 models from 14 studies. The meta-analysis results are shown in Fig. 3. The pooled AUC of the prediction models was 0.84 (95% CI: 0.82, 0.86). Substantial heterogeneity was observed among the studies (I2 = 69.7%, p < 0.001), indicating the need for further subgroup analysis. Forest plot from a meta-analysis of AUC in 35 prediction models Subgroup analysis We performed subgroup analyses to investigate the impact of data type, model features, and prediction outcomes on the diagnostic performance and between-study heterogeneity of the prediction models. Subgroup analysis by data type showed a pooled AUC of 0.85 (95% CI: 0.82, 0.87) with negligible heterogeneity (I2 = 0.0%, p = 0.538) in the test cohort. In the validation cohort, the pooled AUC was 0.83 (95% CI: 0.79, 0.86) with high heterogeneity (I2 = 82.3%, p < 0.001). The difference in heterogeneity between these two cohorts was not statistically significant (p 255 = 0.374). The results are shown in Supplementary Figure S1. Subgroup analysis by prediction outcomes indicated that models predicting pathological complete response (pCR) had a pooled AUC of 0.84 (95% CI: 0.81, 0.88) with high heterogeneity (I2 = 80.0%, p < 0.001). Models predicting LNM showed a pooled AUC of 0.84 259 (95% CI: 0.81, 0.87), also with no significant heterogeneity (I2 = 36.6%, p = 0. 106). Models predicting PR had a pooled AUC of 0.79 (95% CI: 0.71, 0.86), with no significant heterogeneity observed (I2 = 47.3%, p = 0. 128). The test for heterogeneity between subgroups was not statistically significant (p = 0.412). The results are shown in Supplementary Figure S2. Subgroup analysis by model features revealed that models combining ultrasound with clinical features achieved a pooled AUC of 0.87 (95% CI: 0.84, 0.89) with high heterogeneity (I2 = 58.9%, p = 0.001), while ultrasound-only models had a pooled AUC of 0.80 (95% CI: 0.77, 0.84) and also high heterogeneity (I2 = 76.7%, p < 0.001). The test for subgroup differences in heterogeneity was statistically significant (p = 0.009). The results are shown in Fig. 4. Forest plot from meta-analysis: subgroup analysis by model features (US + Clin vs. US) Based on this, combined analysis of model features and prediction outcomes identified a statistically significant interaction (p = 0.026 for subgroup differences). Ultrasound plus clinical features achieved the highest performance when predicting pCR, with an AUC of 0.87 (95% CI: 0.84, 0.91), albeit with high heterogeneity (I2 = 66.5%, p = 0.002). In contrast, ultrasound-only models had the lowest performance in predicting PR, with an AUC of 0.75 (95% CI: 0.67, 0.82) and no observed heterogeneity (I2 = 0.0%). The results are shown in Fig. 5. Forest plot from meta-analysis: combined subgroup analysis by model features and prediction outcomes In the combined analysis of data type and model features within the validation cohort, the combined-feature models yielded an AUC of 0.86 (95% CI: 0.82, 0.91), outperforming ultrasound-only models, which had an AUC of 0.80 (95% CI: 0.75, 0.85). High heterogeneity was present in both model types. In the test cohort, combined-feature models also performed better, with an AUC of 0.86 (95% CI: 0.83, 0.89), compared to ultrasound-only models, which had an AUC of 0.81 (95% CI: 0.77, 0.85). Both model types showed no observed heterogeneity, and the between-subgroup difference was borderline significant (p = 0.058). The results are shown in Supplementary Figure S3. The combined analysis of data type and prediction outcomes showed no statistically significant subgroup differences (p = 0.313), indicating consistent model performance across these subgroups. The results are shown in Supplementary Figure S4. Three-way subgroup analysis revealed no statistically significant interaction among data type, model features, and prediction outcomes (p > 0.05). Nevertheless, descriptive trends indicated performance variations across different combinations. The combined-feature models predicting pCR in the test cohort showed the best performance, with an AUC of 0.88 (95% CI: 0.84, 0.92) and low heterogeneity (I2 = 0.0%). Ultrasound-only models predicting PR in the validation cohort exhibited the poorest performance, with an AUC of 0.75 (95% CI: 0.67, 0.82) and similarly low heterogeneity (I2 = 0.0%). The results are shown in Supplementary Figure S5. We additionally performed a subgroup analysis according to the ultrasound technique used in each model (B-mode alone vs. B-mode + Doppler). The results are presented in Supplementary Figure S6. The B-mode alone group (23 models) showed a pooled AUC of 0.84 (95% CI: 0.81–0.87) with high heterogeneity (I2 = 78.5%). The B-mode + Doppler group 306 (12 models) yielded a pooled AUC of 0.87 (95% CI: 0.79–0.96) with moderate heterogeneity (I2 = 45.3%). The difference between the two subgroups was not statistically significant (p = 0.47), indicating that the addition of Doppler features did not significantly improve model discriminative performance. Given that this analysis was performed post hoc and the Doppler group had a limited sample size, we did not incorporate ultrasound technique into the two-way or three-way interaction analyses. In summary, although the highest-order analysis found no statistically significant interactions, our results suggest that combining ultrasound with clinical features currently represents the most promising and reproducible strategy for predicting pCR in the test cohort. Sensitivity analysis The results of the sensitivity analysis are presented in Fig. 6. In the leave-one-out analysis, the point estimate of the pooled effect size remained stable. It ranged from −0.18 to −0. 15 after removing each study one by one. The 95% CIs after each study was excluded ranged from −0.21 to −0.14, and all CIs consistently included the overall pooled effect size. No single study had a substantial influence on the direction or magnitude of the overall results, indicating that the meta-analysis conclusions are robust and not overly affected by any individual study. Sensitivity analysis of the meta-analysis results Publication bias The funnel plot showed that the effect size estimates were roughly symmetric around the pooled effect size, with most data points within the 95% confidence limits. No evident asymmetry was observed in the visual inspection of the funnel plot. However, statistical tests revealed potential publication bias: Begg's test (z = −2.74, p = 0.006) and Egger's test (slope coefficient = −0.04, p = 0.026; bias coefficient = −2.80, p = 0.000). The results are presented in Fig. 7. Funnel plot analysis for publication bias Discussion Overall diagnostic performance Recently, more studies have focused on prediction models for NAC response in breast cancer. Individual studies indicate that ultrasound-based prediction models have diagnostic value for assessing NAC response [18–31]. However, the overall diagnostic performance, risk of bias, and optimal clinical application scenarios of these models remain unclear. Therefore, a meta-analysis is needed to provide evidence-based support. This study is the first meta-analysis to systematically evaluate ultrasound-based prediction models for NAC response in breast cancer. Our findings demonstrate that these models have good diagnostic accuracy, with a pooled AUC of 0.84 (95% CI: 0.82, 0.86). Implications for radiotherapy planning Additionally, accurate prediction of NAC response may also inform locoregional radiotherapy decisions. For patients who achieve a pathologic complete response (pCR), de-escalation of radiation volumes—such as omitting comprehensive nodal irradiation—has been proposed in ongoing clinical trials [37] (e.g., NSABP B-51/RTOG 1304) and represents a promising strategy to reduce treatment-related morbidity [38]. Conversely, patients with substantial residual disease after NAC, particularly those with residual axillary nodal involvement, are at higher risk of locoregional recurrence [39] and may benefit from post-mastectomy radiotherapy or intensified regional nodal irradiation. Imaging-based early response assessment, including ultrasound-based models, could help stratify patients into these different paradigms: those with a high predicted probability of pCR might be candidates for radiotherapy deescalation trials, while those with predicted residual disease could be considered for escalated locoregional treatment. However, such applications remain investigational and require prospective validation integrating radiologic predictions with post-NAC pathologic findings. Our findings suggest that ultrasound-based models, especially when combined with clinical features, hold preliminary promise for this purpose, but current evidence is not yet sufficient to support practice-changing decisions in radiotherapy planning. Multimodal integration in clinical practice In current clinical practice, ultrasound-based prediction models are unlikely to function as stand-alone decision-making tools. Instead, they could be integrated into a multimodal framework alongside other complementary modalities. For example, when ultrasound suggests incomplete response (high residual tumor burden), MRI may provide additional morphological and functional information (e.g., tumor heterogeneity, perfusion dynamics) to refine assessment [40]. When imaging findings are equivocal—a common scenario in hormone receptor-positive/HER2-negative tumors—post-NAC core needle biopsy of the tumor bed provides definitive pathologic confirmation of residual disease [41]. For triple-negative and HER2-positive breast cancers, where the prognostic value of pCR is strongest, ultrasound predictions could be combined with genomic assays (e.g., residual cancer burden index [42], circulating tumor DNA monitoring [43]) to enhance predictive accuracy. The optimal integration strategy likely involves a risk-adaptive algorithm: for tumors with concordant favorable findings across ultrasound, MRI, and clinical features, close observation may be considered; for discordant or high-risk scenarios, tissue confirmation should remain the standard of care. Prospective studies are needed to define the precise role and added value of ultrasound within such multimodal frameworks. Superiority of combined−feature models According to diagnostic performance criteria, an AUC above 0.80 indicates moderate to good diagnostic capability, suggesting preliminary potential for ultrasound-based models to become exploratory tools in NAC decision-making pending further validation. The 35 prediction models included in this analysis employed varying model features, predictive outcomes, and datasets. Subgroup analysis revealed a significant interaction between model features and the prediction outcomes. Specifically, models combining ultrasound with clinical features for predicting pCR in the test cohort demonstrated the best and most consistent performance among the included studies, with an AUC of 0.88 (95% CI: 0.84, 0.92) and low heterogeneity (I2 = 0.0%). This improved performance is likely because these models integrate imaging and molecular subtyping information, providing a more comprehensive foundation for more accurate predictions [44–46]. Sources of heterogeneity The meta-analysis yielded a pooled AUC of 0.84 (95% CI: 0.82, 0.86), but significant heterogeneity was observed (I2 = 69.7%, p < 0.001). Subgroup analyses identified data type, model features, and prediction outcomes as the primary sources of heterogeneity. Among these, model features were the most critical, showing statistically significant subgroup differences. Models combining radiomic features with clinical data, such as molecular subtypes, showed significantly better diagnostic performance than ultrasound-only models. This strongly suggests that imaging features alone provide limited information [47], and that combining them with clinicopathological features is an effective strategy to enhance model performance [44]. Such findings align with clinical understanding, since chemotherapy response depends not only on tumor morphology and perfusion captured by imaging but also on intrinsic biological characteristics like molecular subtypes [48, 49]. The integration of radiomics and clinical features represents a major trend in medical AI [50]. Supporting evidence comes from MRI-based NAC prediction studies in breast cancer, which also show that combined models outperform pure imaging models [51, 52]. By combining these data types, a more comprehensive characterization of tumor heterogeneity is possible, leading to more accurate chemosensitivity prediction. The heterogeneity analysis based on dataset type showed negligible heterogeneity and stable performance in the test cohort (AUC = 0.85, 95% CI: 0.82, 0.87, I2 = 0.00%), while the validation cohort exhibited high heterogeneity (I2 = 82.3%, p < 0.001). The test cohort represents model generalizability to new data, and its low heterogeneity and high AUC enhance the credibility of the findings. In contrast, the high heterogeneity in the validation cohort likely results from differences in internal validation methods and data splitting ratios across studies. Variations in training-validation cohort splits (e.g., 7:3, 3:1, or 8:2) directly impact the sample size and composition for model training and validation. More importantly, methodological differences in validation existed. Most studies used simple random split internal validation within a single center, which may bias results due to dependence on the specific dataset distribution, while others employed more rigorous approaches. For instance, Liu et al. (2024) [21] used an 8:2 split combined with fivefold cross-validation; Jiang et al. (2021) [31] utilized a completely independent external validation cohort; notably, Gu et al. (2022) [24] employed a 3:1 data split and conducted multi-center external validation. Subgroup analysis by prediction outcomes indicated the highest heterogeneity among studies predicting pCR (I2 = 80.0%). This may be attributed to varying definitions of pCR across studies. Most studies predicting pCR used the ypT0/is ypN0 standard (no invasive carcinoma in breast and axillary lymph nodes, allowing for ductal carcinoma in situ). However, Wang et al. (2023) [28] focused solely on breast pCR and disregarded nodal status. These differing definitions of pathologic response directly influence the prediction target and the model performance, constituting a major source of heterogeneity [53]. In addition, heterogeneity in ultrasound techniques (B-mode alone vs. B-mode + Doppler) may have contributed to the observed overall heterogeneity. However, our subgroup analysis did not detect a significant performance difference between these two technique groups (p = 0.47), suggesting that technique diversity alone does not explain the substantial heterogeneity (I2 = 69.7%). Other factors, such as differences in study design, patient populations, and model development pipelines, likely play a larger role. Additionally, the included molecular subtypes varied among studies. Specifically, most studies included the four major molecular subtypes, but two studies only provided histological types, and one did not specify the pathological type. Research indicates that although the Collaborative Trials in Neoadjuvant Breast Cancer (CTNeoBC) consensus recommends ypT0/is ypN0 as the standard pCR definition and endpoint [54], the clinical significance of this definition varies significantly across molecular subtypes [55]. For example, pCR is more strongly associated with long-term survival benefits in HER2-positive and triple-negative breast cancers [54, 56, 57]. Therefore, future studies should not only explicitly report the pCR definition used but also consider stratified modeling or analysis by molecular subtype to provide more targeted predictions. In contrast, studies predicting LNM showed lower heterogeneity (I2 = 36.6%, p = 0. 106), likely because the definition is more standardized and objective. Risk of bias assessment PROBAST assessment rated only two of the 14 included studies as low risk of bias, with most high risks arising from the Participants (Domain 1) and Analysis (Domain 4) domains. Regarding participant selection (Domain 1), all 14 studies were retrospective, and only two were multi-center, which introduced selection bias. Furthermore, many studies commonly excluded patients with multifocal lesions or poor image quality. This led to insufficient representativeness of the study samples and severely limited the models' generalizability. Concerning the analysis domain (Domain 4), risks primarily arose from insufficient sample size and overfitting. Some studies (e.g., Liu et al. and Zhang et al.) employed dimensionality reduction methods like Least Absolute Shrinkage and Selection Operator (LASSO) regression and incorporated internal validation, which resulted in an 'Unclear' risk rating; however, their sample sizes remained limited. Other studies such as Wang et al., were rated high risk of bias due to a small training cohort sample size relative to the numbers of features in the final model, coupled with a lack of external validation and insufficient model calibration. These findings suggest that the reported high performance metrics resulted from overfitting to specific datasets rather than true predictive ability, leading to an overestimation of the models' diagnostic performance; therefore, their real-world performance should be interpreted with caution. Publication bias and generalizability challenges Sensitivity analysis confirmed the stability of the findings. Although the funnel plot appeared symmetrical, Begg's and Egger's tests indicated potential publication bias. These results, consistent with the quality assessment, suggest a risk of overestimation of the models' diagnostic performance and thus call for cautious interpretation of the meta-analysis results. High risk of bias and publication bias are not merely limitations of this meta-analysis but represent common challenges across the AI-assisted diagnostics field, where models often perform well on internal data but fail to generalize to external real-world settings [58]. Future high-quality studies will be needed to provide more robust evidence-based support. Comparison with previously published meta−analyses Several meta-analyses [32–35] have previously evaluated ultrasound-based prediction of NAC response in breast cancer. Shi et al. (2026) broadly evaluated six conventional ultrasound modalities across 72 studies, concluding that ultrasound “is helpful”. Valizadeh et al. (2025) focused on machine learning and radiomics (22 studies, search up to August 2023), concluding “strong potential” and highlighting delta radiomics. Li et al. (2024) examined ultrasound-based radiomics (8 studies, search up to January 2023) and reported pooled sensitivity of 87% and specificity of 78%. Jin et al. (2021) specifically assessed strain and shear wave elastography (15 studies), reporting AUCs of 0.89 and 0.82, respectively. In contrast, our study offers several distinct contributions. First, we focus specifically on radiomics and machine learning models with an updated search (May 2025), including 35 models from 14 studies –providing a more contemporary synthesis than previous reviews. Second, we used PROBAST (designed for prediction models) rather than QUADAS-2 or METRICS, revealing that 71.4% of included studies were at high overall risk of bias – a critical finding not emphasized in any prior meta-analysis. Third, we directly compared ultrasound-only vs. combined-feature models, demonstrating that adding clinical features significantly improves performance (pooled AUC 0.87 vs. 0.80, p = 0.009) –a direct quantitative comparison of the discriminative performance (pooled AUC) between ultrasound-only and combined-feature models has not been performed by any previous meta-analysis. Fourth, while previous meta-analyses included studies from multiple countries, none explicitly discussed how regional differences in breast cancer epidemiology, molecular subtype distribution, NAC protocols, and imaging standards may limit generalizability. Our study explicitly acknowledges this as a major limitation. Fifth, we offer a tempered, hypothesis-generating conclusion rather than ‟ strong potential” or ‟ helpful”. Finally, we performed a post-hoc subgroup analysis by ultrasound technique (B-mode alone vs. B−mode+Doppler, p = 0.47), which has not been reported elsewhere. Together, these differences position our study as a critical methodological appraisal that complements and extends prior work. Delta radiomics Two studies in our meta-analysis (Yang et al., 2022 [29]; Liu et al., 2024 [21]) explicitly reported delta radiomics features – defined as changes in radiomic features between pre- treatment and early post- treatment (after 2 cycles) ultrasound. Yang et al. found that their delta radiomics model did not outperform the post- treatment radiomics model (AUC 0.743 vs. 0.811), while Liu et al. reported comparable performance (AUC 0.850 vs. 0.863). However, with only two studies, a meta- analytic synthesis was not feasible. Valizadeh et al. (2025) [33]– published after our search cutoff – reported that delta radiomics may enhance sensitivity for complete response prediction. Our data neither confirm nor contradict this due to the limited number of primary studies. We therefore call for future prospective studies with standardized longitudinal imaging protocols to rigorously evaluate the incremental value of delta radiomics over baseline or post- treatment radiomics. Limitations This study had several limitations. First, all 14 included studies were retrospective. Crucially, all were conducted in China, which severely limits the generalizability of the findings to other populations and healthcare settings. Furthermore, geographic generalizability is further constrained by regional differences in breast cancer epidemiology, molecular subtype distribution, neoadjuvant chemotherapy regimens, and ultrasound imaging standards [59]. For instance, Asian populations have a higher prevalence of luminal subtypes and a lower prevalence of triple-negative breast cancer compared with Caucasian populations [60]. Additionally, neoadjuvant chemotherapy protocols (e.g., drug selection, cycle length) [4] and ultrasound image acquisition parameters may vary substantially across countries. Consequently, our pooled AUC estimate of 0.84 may not directly apply to North American or European settings, where different patient demographics and treatment practices prevail. External validation of these ultrasound-based models in geographically diverse, multicenter cohorts is essential before any clinical implementation can be considered. Second, variations in specific algorithms and ultrasound image preprocessing pipelines among the included models were potential sources of heterogeneity that could not be precisely measured or analyzed. Third, the included studies varied in ultrasound techniques (B-mode alone vs. B-mode + Doppler). Although we performed a subgroup analysis comparing the two and found no significant difference in pooled AUC, the limited number of studies using Doppler (n = 12 models) precludes a definitive conclusion. Future studies with larger sample sizes and standardized technical protocols, potentially incorporating other advanced ultrasound modalities, are needed to further explore this source of heterogeneity. Although extensive subgroup analyses were conducted, some potential confounding factors cannot be entirely ruled out. Finally, only two of the 14 included studies reported delta radiomics features, precluding a robust subgroup meta-analysis; the potential added value of delta radiomics remains to be validated in larger, standardized cohorts. Future directions The publication of multiple overlapping systematic reviews on ultrasound- based prediction of NAC response – including Shi et al. (2026), Valizadeh et al. (2025), and the present study – within a short timeframe highlights the need for more efficient knowledge synthesis. Rather than simply duplicating efforts, future research should prioritize: (1) prospective, multicenter validation studies with standardized imaging protocols and outcome definitions; (2) individual participant data meta- analyses, which would allow for nuanced subgroup analyses (e.g., by molecular subtype, ultrasound modality, geographic region) that are not possible with aggregate data; and (3) living systematic reviews that are updated continuously as new evidence emerges, thereby preventing research waste and ensuring that clinical decision-making reflects the most current evidence. Additionally, future meta-analyses should move beyond simply pooling AUCs and provide rigorous risk- of-bias assessments using PROBAST, explicitly discuss geographic generalizability, and offer tempered conclusions that acknowledge the current evidence as exploratory and hypothesis-generating rather than practice-changing. Given the high risk of bias, substantial heterogeneity, and geographic limitations identified in this meta-analysis, the current evidence should be considered exploratory and hypothesis-generating rather than practice-changing. Future research should focus on three key areas. First, prospective multi-center cohort studies with pre-registered protocols are essential to fundamentally reduce participant selection bias. Second, during statistical analysis, researchers should ensure adequate sample sizes and employ more rigorous validation strategies, particularly by using completely independent test cohorts for final performance evaluation. Finally, researchers should strictly adhere to international reporting guidelines such as Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) to ensure complete and transparent disclosure of all model parameters and development details. These systematic improvements will significantly reduce the risk of bias. As a result, higher-quality prediction models with greater clinical utility can be developed. Conclusion Ultrasound-based prediction models show encouraging discriminative ability for predicting NAC efficacy in breast cancer, particularly models integrating clinical features for predicting pCR, which demonstrate moderate to good performance. However, the current evidence is characterized by a high risk of bias and potential publication bias, and a narrow geographic origin (all from China), which substantially limits the external validity of the findings. Therefore, the current evidence should be considered exploratory and hypothesis-generating rather than practice-changing. Supplementary Information Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Supplementary Material 4. Supplementary Material 5. Supplementary Material 6. Supplementary Material 7. Acknowledgements Not applicable. We extend our appreciation to all researchers whose published work was included in this meta-analysis. Abbreviation Area Under the Curve Breast Conserving Surgery Confidence Interval Collaborative Trials in Neoadjuvant Breast Cancer False-Negative Rate Least Absolute Shrinkage and Selection Operator Lymph Node Metastasis Neoadjuvant Chemotherapy Pathological Complete Response Population, Intervention, Comparator, Outcome, Timing, Setting Partial Response Prediction Model Risk of Bias Assessment Tool Response Evaluation Criteria in Solid Tumors Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis Authors’ contributions Conceptualization: Cuo Yi, Xi Yang. Methodology: Cuo Yi, Qiang Zhang. Formal analysis: Cuo Yi, Qiang Zhang. Investigation: Ting Xu, Rui You. Data curation: Ting Xu, Rui You. Resources: Qiang Zhang. Writing – original draft: Cuo Yi. Writing – review & editing: Ruihan Liu, Xi Yang. Visualization: Cuo Yi. Supervision: Ruihan Liu, Xi Yang. Project administration: Ruihan Liu. Funding acquisition: Xi Yang. All authors read and approved the final manuscript. Funding This study was funded by the 2026 National Natural Science Foundation of China (Grant No. 82572258) and 2026 Chongqing Science and Health Joint Medical Research Project (Grant No. 2026MSXM135). Data availability All data generated or analyzed during this meta-analysis are included in this published article and its supplementary information files. The datasets supporting the conclusions of this article are derived from previously published studies, which are listed in the reference section. Access to the original datasets from these studies should be sought from the corresponding authors of the respective primary studies. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. This meta-analysis is a secondary study based on aggregated data from. previously published articles; no individual patient data or identifiable images are involved. Competing interest The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Ruihan Liu, Email: ruihan_l@163.com. Xi Yang, Email: yangxi210@cqu.edu.cn. References Associated Data Supplementary Materials Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Supplementary Material 4. Supplementary Material 5. Supplementary Material 6. Supplementary Material 7. Data Availability Statement All data generated or analyzed during this meta-analysis are included in this published article and its supplementary information files. The datasets supporting the conclusions of this article are derived from previously published studies, which are listed in the reference section. Access to the original datasets from these studies should be sought from the corresponding authors of the respective primary studies.