Predicting response to patients with gastric cancer via dynamic-aware model with longitudinal liquid biopsy data

Predicting response to patients with gastric cancer via dynamic-aware model with longitudinal liquid biopsy data | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Predicting response to patients with gastric cancer via dynamic-aware model with longitudinal liquid biopsy data zifan chen, Jie Zhao, Yanyan Li, Yilin Li, Xinyu Nan, Huimin Liu, and 5 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5181858/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Gastric cancer (GC) presents challenges in predicting treatment responses due to patient-specific heterogeneity. Recently, liquid biopsies have emerged as a valuable data modality, providing essential cellular and molecular insights and facilitating the capture of time-sensitive information. This study aimed to harness artificial intelligence (AI) technology to analyze longitudinal liquid biopsy data. We collected a dataset from longitudinal liquid biopsies of 91 patients at Peking Cancer Hospital, spanning from July 2019 to April 2022, including 1,895 tumor-related cellular images and 1,698 tumor marker indices. Subsequently, we introduced a Dynamic-Aware Model (DAM) to predict GC treatment responses. DAM incorporates dynamic data through AI components for in-depth longitudinal analysis. Using three-fold cross-validation, DAM exhibited superior performance in predicting treatment responses compared to traditional methods (AUCs: 0.807 vs. 0.582), maintained stable efficacy in the test set (AUC: 0.802), and accurately predicted responses from early treatment data. Moreover, DAM's visual analysis of attention mechanisms identified six key visual features associated strongly with treatment responses. These findings represent a pioneering effort in applying AI technology for interpreting longitudinal liquid biopsy data and employ visual analytics in GC, offering a promising avenue toward precise response prediction and tailored treatment strategies for patients with GC. Figures Figure 1 Figure 2 Figure 3 Figure 4 1. Background Gastric cancer (GC), ranking fifth globally in prevalence, stands as the third leading cause of cancer-related mortality 1 , 2 . The widespread occurrence of GC demands the formulation of effective treatments and precise predictive models for treatment responses 3 , 4 . However, GC treatment is challenged by patient heterogeneity, hindering accurate treatment response predictions 5 , 6 . Such heterogeneity, rooted in cellular characteristics and morphological variations, is critical in devising patient-specific treatment plans 7 , 8 . In response, recent advancements in analyzing liquid biopsy data, including circulating tumor cells (CTCs), circulating endothelial cells (CECs), and tumor markers, have emerged as a potent tool in cancer treatment management 9 , 10 . Analysis of tumor-related cells and tumor markers in liquid biopsies offers dual advantages: firstly, it acts as a cellular-level biomarker, providing insights into cellular heterogeneity and its association with treatment response 11 – 14 ; secondly, liquid biopsies are more convenient than traditional diagnostics like computerized tomography (CT) scans and histopathological examinations 11 , 15 , facilitating the gathering of time-sensitive data essential for understanding a patient’s evolving cellular changes during treatment 12 . Indeed, longitudinal liquid biopsy data is expected to provide personalized treatment strategies to adapt to the changing nature of the disease and patient response. Recent advancements in cancer diagnosis and treatment management have largely benefited from cellular biomarker applications, including CTCs 16 – 19 , CECs 20 , 21 , and tumor markers 22 , 23 . These developments have substantially improved our understanding of cancer, from early detection 20 , 24 – 26 to treatment outcome prediction 17 , 18 , 21 , 27 , 28 . However, many previous studies 29 – 31 have relied on cell-counting-based statistical approaches for outcome assessment. While useful, these methods frequently overlook potential insights that longitudinal liquid biopsy data can provide. Conversely, artificial intelligence (AI), particularly in sequence modeling 32 – 34 and visual information extraction 35 , 36 , shows promise in revolutionizing biomedical data analysis 37 – 41 . AI's ability to process large datasets and decode complex patterns may lead to more precise and personalized treatment efficacy assessments 42 – 44 . Integrating AI into GC management enables the creation of uniquely tailored treatment strategies for each patient's molecular profile, enhancing treatment efficacy and minimizing invasiveness 45 – 47 . Although some studies 48 – 50 have begun exploring the role of dynamic data in cancer treatment management, to our knowledge, none have yet investigated longitudinal liquid biopsies, especially regarding dynamic tumor-related cellular images and dynamic tumor markers. This highlights the need for more advanced analytical methods in longitudinal liquid biopsy data analysis. Addressing these challenges, we compiled a comprehensive longitudinal dataset from 91 patients with GC treated at Peking Cancer Hospital between July 2019 and April 2022. This dataset includes 1,895 aneuploid tumor-related cellular images and 1,698 tumor marker indices from six markers, derived from longitudinal liquid biopsies collected across all available follow-up visits. We randomly divided the dataset into a training set consisting of 74 patients, using three-fold cross-validation for model development, and a test set of 17 patients to assess the model’s generalizability and robustness. We developed a deep-learning-based dynamic-aware model (DAM) to precisely predict GC treatment responses. DAM uniquely tackles the challenge of interpreting complex patterns and temporal dynamics in treatment response predictions by combining convolutional and fully connected neural networks for feature extraction with attention mechanisms for information integration. Specifically, it uses self-attention-based modules to integrate multi-object and multi-temporal data, and a cross-attention-based module to merge mismatched multisource dynamic data. Experimental results confirm DAM’s efficacy in deriving insights from longitudinal liquid biopsies and accurately predicting treatment responses. Furthermore, we identified six dynamic focus area features via DAM’s visual analysis and conducted preliminary studies to assess its potential for interpreting liquid biopsy visual data. 2. Methods 2.1. Ethics The Peking University Cancer Hospital Ethics Committee granted ethical approval for this study (approval number: 2020KT08). All participants or their legally authorized representatives provided informed consent. 2.2. Patients and data collection The study included patients with GC from July 2019 to April 2022. For each patient, longitudinal liquid biopsies were collected, comprising dynamic tumor-related aneuploid cellular images (CTCs and CECs) and dynamic tumor marker indices, at baseline and during subsequent follow-ups (Figs. 1 A and S1). Blood samples underwent density gradient centrifugation and microfluidic isolation to enrich CTCs and CECs. Subsequently, isolated cells were fixed onto slides, and stained with specific markers, including cluster of differentiation 31 (CD31), cluster of differentiation 45 (CD45), centromere protein 8 (CEP8), and 4',6-diamidino-2-phenylindole (DAPI), using specific iFISH (immunostaining-FISH) staining following the manufacturer’s protocol 51 (Cytelligen, San Diego, CA, USA) with minor alteration, and then imaged using the automated Metafer-i•FISH® CTC 3D scanning and image analysis system 52 codeveloped by Carl Zeiss (Oberkochen, Germany), MetaSystems (Altlussheim, Germany) and Cytelligen to capture high-resolution multi-channel overlay images for analysis (More details in Text S1). Furthermore, blood samples were analyzed in the laboratory to measure levels of various tumor markers, including alpha-fetoprotein (AFP), carcinoembryonic antigen (CEA), carbohydrate antigen 19 − 9 (CA19-9), cancer antigen 72 − 4 (CA72-4), cancer antigen (CA125), and neuron-specific enolase (NSE), constituting the collected tumor marker indices. In patients with GC, the definition of treatment response is as follows: Patients who achieve a complete response (CR) or partial response (PR) according to the Response Evaluation Criteria in Solid Tumors (RECIST) are categorized as responders. Conversely, those exhibiting progressive disease (PD) or stable disease (SD) are classified as non-responders. 2.3. The architecture of DAM We introduced the DAM, a deep-learning-based dynamic-aware model designed to analyze dynamic tumor-related cellular images and dynamic tumor marker indices from longitudinal liquid biopsies to predict GC treatment responses. DAM’s architecture, shown in Fig. 1 B, comprises a comprehensive system with five integrated components. DAM begins with the cellular aggregator, employing a ResNet-18 35 , as the feature extractor and a dual-stage Transformer 34 , to create a unified representation of cellular data across various time points (Fig. S2 ). Concurrently, the tumor marker aggregator utilizes a dual-layer perceptron and a similar dual-stage Transformer for the integration of tumor marker indices over time (Fig. S3 ). Subsequently, the temporal interaction module (TIM) aligns these dynamic features using an advanced cross-attention mechanism (Fig. S4 ), ensuring effective synchronization and integration of mismatched multisource temporal data. Following this, the temporal aggregator employs a quad-stage Transformer to consolidate data from various time points into a comprehensive patient-centric feature representation (Fig. S5). Lastly, the predictor component utilizes a three-layer multilayer perceptron (MLP) to classify patients into responder or non-responder categories based on the integrated features (Fig. S6). All of these components operate in concert to ensure a robust and accurate analysis of longitudinal liquid biopsies. 2.4. Training and assessment procedures of DAM This study meticulously designed experiments to ensure DAM’s effective training and fair verification using a constrained dataset (Fig. 1 C). The dataset was randomly divided, allocating 74 patients for training and a separate set of 17 patients reserved for independent testing. The training dataset was further subdivided into three subsets: fold-1 (F1), fold-2 (F2), and fold-3 (F3), comprising 26, 26, and 22 patients with GC, respectively. A three-fold cross-validation strategy 53 was implemented, rotating these subsets between training and validation to fine-tune the model's hyperparameters and architecture. This process yielded three distinct trained models, collectively forming an ensemble model evaluated against the independent test set for robustness. To enhance the model's applicability in real-world medical scenarios, we introduced a dynamic longitudinal elimination strategy during training (Fig. S7). This strategy entailed an 80% non-repetitive random sampling along the temporal dimension of the dynamic tumor-related cellular images and dynamic tumor markers, effectively mimicking a data-level dropout. This strategy fosters model resilience and adaptability to practical challenges like missing some temporal data. When validation and testing, we utilized all available longitudinal data without any random sampling to ensure a thorough assessment. After training, we froze model parameters for further analysis. We first applied softmax activation function to predictive probability of DAM and used them to plot receiver operating characteristic (ROC) curves and calculate an area under the curve (AUC) scores. Furthermore, employing the GradCAM algorithm 54 on the final convolutional layer of the cellular aggregator produced insightful attention maps. Based on these maps, six related dynamic features were quantified (Fig. 4 A). Specifically, for a particular patient with a total of \(\:N\) tumor-related cellular images \(\:\{{I}_{1},{I}_{2},\cdots\:,{I}_{N}\}\) across \(\:T\:\) time points, each image is associated with a corresponding attention map generated by GradCAM, denoted as \(\:\{{A}_{1},{A}_{2},\cdots\:,{A}_{N}\}\) . These attention maps were then subjected to binary segmentation using a 75-th percentile threshold to delineate focus areas \(\:\{{F}_{1},{F}_{2},\cdots\:,{F}_{N}\}\) , where the focus area is defined as the total area identified as foreground in the binary segmentation. The dynamic features, namely, focus area variability (VarFA), minimum focus area (MinFA), maximum focus area (MaxFA), average focus area (AvgFA), and median focus area (MedFA), were denoted as follows, $$\:\text{A}\text{v}\text{g}\text{F}\text{A}=\frac{1}{N}{\sum\:}_{i=1}^{N}{F}_{i}^{\:},$$ $$\:\text{V}\text{a}\text{r}\text{F}\text{A}=\frac{1}{N-1}{\sum\:}_{i=1}^{N}{\left({F}_{i}-AvgFA\right)}^{2},$$ $$\:\text{M}\text{i}\text{n}\text{F}\text{A}=\text{m}\text{i}\text{n}\{{F}_{1},{F}_{2},\cdots\:,{F}_{N}\},$$ $$\:\text{M}\text{a}\text{x}\text{F}\text{A}=\text{m}\text{a}\text{x}\{{F}_{1},{F}_{2},\cdots\:,{F}_{N}\},$$ $$\:\text{M}\text{e}\text{d}\text{F}\text{A}=\text{m}\text{e}\text{d}\text{i}\text{a}\text{n}\left\{{F}_{1},{F}_{2},\cdots\:,{F}_{N}\right\}.$$ Additionally, focus area dispersion (DisFA) was denoted by determining the number of connected domains within the regions of interest, expressed as $$\:\text{D}\text{i}\text{s}\text{F}\text{A}={\sum\:}_{i=1}^{N}{D}_{i},$$ where \(\:{D}_{i}\) represents the number of connected domains in the \(\:i\) -th image. Finally, a three-layer MLP, was trained using these features to uncover nonlinear relationships between the features and treatment responses. 2.5. Statistical analyses The sample size determination was contingent upon the number of patients fulfilling the inclusion criteria (first- or second-line treatment with at least two time-point data) rather than a pre-established statistical methodology. The allocation of subjects across different groups, as delineated in Figs. 2 C–D, 4 A, and 4 C, was evaluated utilizing the Mann-Whitney U test. We conducted our statistical analyses via R software (version 4.1.3) or Python (version 3.7.10). A P -value threshold below 0.05 was designated as the criterion for statistical significance. To ensure the reproducibility of this study, comprehensive methodologies and data management protocols are thoroughly documented in the supplementary information (Figs. S1–S10 and Texts S1–S3). The robustness of DAM was augmented through a three-fold cross-validation strategy and was further validated by an independent test set. Moreover, the source code, implemented by PyTorch 55 , is available in the supplementary materials. 2.6. Role of funders The funders played no role in the study design, data collection, data analysis, data interpretation, and writing of the report. 3. Results 3.1. Patient profile and data characteristics The data in this study were collected from patients with GC treated at Peking Cancer Hospital during the period between July 2019 and April 2022. The dataset comprises longitudinal liquid biopsy data from 91 patients, encompassing 1,895 aneuploid tumor-related cellular images and 1,698 tumor marker indices. The dataset was randomly divided into two a training set of 74 patients for deep-learning model development, using three-fold cross-validation, and a test set of the remaining 17 patients to assess the model's generalizability and robustness (Table 1 ). The median age of the patients was 65 years, with an interquartile range of 57 to 72 years. Males constituted 80.22% of the dataset, indicating a pronounced male predominance. Treatment lines included first-line therapies (80.22%) and other-line therapies (19.78%). Most patients were diagnosed with advanced-stage disease. Table 1 Pathological characteristics of enrolled patients. Patient baseline characteristics N (%) P Fold-1 Fold-2 Fold-3 Test Age ≥60 12 (46.2) 17 (65.4) 12 (54.5) 11 (64.7) 0.482 <60 14 (53.8) 9 (34.6) 10 (45.5) 6 (35.3) Sex Male 22 (84.6) 20 (76.9) 20 (90.9) 11 (64.7) 0.199 Female 4 (15.4) 6 (23.1) 2 (9.1) 6 (35.3) Location GEJ 9 (34.6) 8 (30.8) 7 (31.8) 5 (29.4) 0.985 Non-GEJ 17 (65.4) 18 (69.2) 15 (68.2) 12 (70.6) Differentiation High 0 (0.0) 1 (3.8) 1 (4.5) 0 (0.0) 0.562 Moderate-high 0 (0.0) 0 1 (4.5) 0 (0.0) Moderate 11 (42.3) 13 (50.0) 6 (27.3) 6 (35.3) Moderate-poor 7 (26.9) 5 (19.2) 6 (27.3) 2 (11.8) Poor 8 (30.8) 7 (26.9) 7 (31.8) 9 (52.9) Unknown 0 (0.0) 0 (0.0) 1 (4.5) 0 (0.0) Lauren classification Intestinal type 17 (65.4) 17 (65.4) 14 (63.6) 13 (76.5) 0.818 Diffused type 4 (15.4) 2 (7.7) 3 (13.6) 3 (17.6) Mixed type 4 (15.4) 4 (15.4) 4 (18.2) 1 (5.9) Unknown 1 (3.8) 3 (11.5) 1 (4.5) 0 (0.0) HER2 expression (IHC) Positive 17 (65.4) 18 (69.2) 17 (77.3) 13 (76.5) 0.775 Negative 9 (34.6) 8 (30.8) 5 (22.7) 4 (23.5) Therapeutic efficacy CR 0 (0.0) 1 (3.8) 0 (0.0) 0 (0.0) 0.568 PR 17 (65.4) 13 (50.0) 15 (68.2) 8 (47.1) SD 4 (15.4) 7 (26.9) 6 (27.3) 7 (41,2) PD 5 (19.2) 5 (19.2) 1 (4.5) 2 (11.8) Note: * Abbreviations : GEJ: Esophagogastric junction; IHC: Immunohistochemical; CR: Complete response; PR: Partial response; SD: Stable disease; PD: Progressive disease. A. Bubble diagram shows the heterogeneity in tumor-related cellular images. Each column corresponds to an individual patient, while each row represents different time intervals, including Pre-Baseline (Pre-BL), within 45 days post-baseline ( \(\:\le\:\) 45d), within 3 months post-baseline ( \(\:\le\:\) 3m), within 6 months post-baseline ( \(\:\le\:\) 6m), within a year post-baseline ( \(\:\le\:\) 1y), and beyond a year post-baseline ( \(\:>\) 1y). The presence of each dot signifies the number of time points collected from a patient within these respective time intervals, with larger dots indicating a higher frequency of sampling. Bar graph adjacent to the right side of the panel further elucidates the frequency of sampling across these distinct time intervals. B. Bubble diagram similarly depicts the heterogeneity in the distribution of tumor marker indices, employing an interpretation approach analogous to that used in A . C. Box plot highlights the heterogeneity in cell counts across the aforementioned time intervals. D. Box plot differentiates the heterogeneity in cell counts across time intervals between responders and non-responders. E. Diagram illustrates the division of these time intervals. The study meticulously tracked the temporal distribution of liquid biopsy metrics to capture the dynamic aspects of GC progression. Figures 2 A– 2 B show the dynamic distribution of tumor-related cellular images and tumor marker indices across patients and time intervals, respectively. The median number of collection time points per patient was three for both tumor-related cellular images and tumor marker indices, with an interquartile range of 1 to 4. This median indicates robust longitudinal data collection, providing a solid foundation for disease progression analysis. Additionally, Fig. 2 C highlights fluctuations in tumor-related cell counts over different time intervals (Fig. 2 E), revealing consistent distribution patterns across all intervals. This consistency, observed regardless of treatment response (Fig. 2 D), underscores the nuanced complexity of disease trajectories reflected in liquid biopsy profiles. 3.2. Enhanced response prediction using the dynamic-aware model compared to the cell-counting-based model Comprehensive analysis shows that the proposed DAM outperforms the traditional cell-counting-based model in predicting GC treatment responses (Fig. 3 ). The ROCs from three-fold cross-validation shows that fold-1 achieved an AUC of 0.739 (95% Confidence Interval [CI]: 0.528–0.922), fold-2 an AUC of 0.845 (95% CI: 0.655–0.976), and fold-3 an AUC of 0.838 (95% CI: 0.639–0.979) (Fig. 3 D). The overall average AUC inclusive of the standard deviation was 0.807 ± 0.048. Furthermore, applied to an independent test set, DAM demonstrated robust performance with an AUC of 0.802 (95% CI: 0.532–1.000; Fig. 3 E). By comparison, the cell-counting-based model, using the XGBoost algorithm based on baseline tumor-related cell counts recorded an average AUC of 0.582 ± 0.037 (Fig. 3 A and Table S1 ). To examine the impact of two longitudinal data sources, including dynamic tumor-related cellular images and dynamic tumor markers, we developed two DAM variants: DAM-TCI and DAM-TM. DAM-TCI focuses on solely utilizing dynamic tumor-related cellular images to predict response; while DAM-TM focuses on solely utilizing dynamic tumor markers. In our experiments, DAM-TCI achieved an average AUC of 0.731 ± 0.047 (Fig. 3 B), whereas DAM-TM achieved an average AUC of 0.725 ± 0.058 (Fig. 3 C). Both variants outperformed the cell-counting-based model, highlighting DAM’s dynamic design for synthesizing dynamic data from different time points. The mismatched collection times for dynamic tumor-related cellular images and tumor marker indices (Fig. S1 ) necessitate the model’s ability to adaptively integrate mismatched multisource dynamic data. We proposed a enhanced cross-attention module (TIM) to deal with this challenge. Integrating dynamic tumor-related cellular images with dynamic tumor marker indices via TIM enhanced performance to 0.807 (Fig. 3 D), surpassing the individual dynamic data sources. This outcome underscores the efficacy of TIM in amalgamating multiple dynamic data types and the complementary role of tumor marker information in tumor-related cellular images. Subsequently, we performed a time-series dynamic analysis of DAM, allowing the model to access data from specified time intervals, including forty-five days (45d), three months (3m), six months (6m), one year (1y), and all available data (all) in the collection (Fig. 2 E). DAM’s performance gradually improved with extended time intervals, showing average AUCs of gradual improvement in DAM’s performance with the extension of the time interval (Fig. 3 F), showing average AUCs of 0.644 (45d), 0.718 (3m), 0.786 (6m), 0.786 (1y), and 0.807 (all). This indicates that including more dynamic information from extended treatment records improves prediction accuracy. Notably, DAM showed modest improvement after six months (an absolute increase of 0.021, from 0.786 to 0.807), in contrast to a big improvement within the first six months (an absolute increase of 0.142, from 0.644 to 0.786). This underscores DAM’s effectiveness in using early treatment data for precise treatment response prediction, aligning with the clinical need to identify non-responders promptly in the treatment process. 3.3. Visual analysis of DAM The previously cell-counting-based approach overlooks the visual characteristics of tumor-related cellular images. In contrast, the proposed DAM enhances complex visual data interpretation, enabling a more comprehensive image analysis. We explored DAM’s visual focus mechanism by analyzing important score distributions for cellular features within the aggregator, categorized by CTCs and CECs (Fig. 4 A). CTCs had significantly higher important scores than CECs (two-sided Mann-Whitney U test, P < 0.0001), emphasizing the crucial role of CTCs in DAM. Consequently, we visualized tumor-related cellular images at the 25-th, 50-th, and 75-th percentile of the important score distribution. Our intuitive observations suggest that larger and more semantically complex CTCs have a more substantial impact on outcomes. This leads us to believe that DAM primarily interprets liquid biopsy visual information through two key visual signals: size and heterogeneity within the image. Furthermore, utilizing the GradCAM algorithm, we visualized DAM’s visual encoder, noting variations in focus area sizes and dispersion across different cellular images (Fig. 4 B). Building on DAM’s visual focus analysis insights, we further explored dynamic focus-area-related visual features (more details in the methods section), including VarFA, MinFA, MaxFA, AvgFA, MedFA, and DisFA (Fig. 4 C). We employed a nonlinear analysis via a three-layer MLP to predict treatment response using these dynamic visual features (Fig. S6), yielding AUC scores of 0.712 (95% CI: 0.444–0.948) in fold-1, 0.804 (95% CI: 0.607–0.955) in fold-2, and 0.819 (95% CI: 0.614–0.983) in fold-3, with an average of 0.778 ± 0.047 (Fig. 4 C), with a comparable performance to DAM on the test set (0.806 vs. 0.802; Fig. 4 D). Moreover, a permutation importance analysis highlighted AvgFA as particularly important, underscoring tumor cell size’s importance in DAM. These analyses offer initial insights into the visual characteristics of tumor-related cellular images, suggesting deep-learning models like DAM have the potential to summarize generalized spatial visual information. 4. Discussion GC remains a significant global health challenge, underscoring the need for precise and early evaluation of treatment effectiveness. Recent studies on liquid biopsy data, including CTCS, CECs, and tumor markers, have highlighted its potential in cancer treatment management. Analyzing liquid biopsy data offers two main advantages: providing crucial cellular-level biomarkers and being more convenient than traditional methods like CT scans and histopathology. Such convenience aids in the acquisition of time-sensitive data. However, a notable research gap exists in processing and interpreting complex, longitudinal liquid biopsy data, highlighting the need for more advanced analytical methods. Our study is the first to apply longitudinal liquid biopsy data in developing an innovative AI model, DAM. DAM utilizes cutting-edge AI technology for integrating multi-source, multi-object, and multi-temporal data and makes accurate response predictions, signifying a considerable advancement in the field. A key finding, distinguishing our research from previous liquid biopsy studies, is DAM’s superior performance in modeling longitudinal liquid biopsies for treatment response predictions. Unlike previously cell-counting-based methods, DAM processes dynamic data from CTCs, CECs, tumor markers, and associated visual imagery concurrently. The comparative analysis underscores DAM’s enhanced capability in prediction response in GC treatment, and an evaluation in the independent test set reveals its robustness. Additionally, two DAM variants, DAM-TCI and DAM-TM, underscore the importance of longitudinal information and demonstrate dynamic modeling’s effectiveness in DAM. The analysis of the impact of dynamic data from diverse treatment time intervals on efficacy prediction reveals that DAM can accurately predict outcomes using early treatment data, aligning closely with clinical usage scenarios. This study also showcases DAM’s module design’s exceptional flexibility and expansiveness, adeptly aggregating diverse elements like tumor-related cellular images. Additionally, the TIM within DAM demonstrates exceptional proficiency in aligning temporally mismatched data from diverse dynamic data sources. By utilizing tumor markers to enhance the analysis of tumor-related cellular images in this study, TIM establishes a groundbreaking approach for data integration. This method is particularly suited for medical contexts, where temporal discrepancies between datasets are common. Its applicability to future scenarios suggests a broad utility beyond the immediate context, offering a promising avenue for integrating diverse data types across a wide range of fields. Besides, within DAM, the temporal aggregator skillfully handles a varied temporal distribution of dynamic data. This innovative design not only accommodates multiple objects, times, and modalities but also seamlessly incorporates non-dynamic (single-time-point) data, such as historical radiological and pathological records. The versatility and scalability inherent in the module design are crucial, facilitating the adaptation and integration of a wide spectrum of data types and sources, thereby enhancing the model's applicability and robustness in complex clinical scenarios. By leveraging advanced deep learning technologies like convolutional neural networks (CNNs) and Transformers, DAM effectively models visual information of tumor-related cellular images, a facet overlooked by previously cell-counting-based methods. This approach facilitated a preliminary analysis of DAM's visual attention patterns. Through detailed evaluations of areas of interest of DAM within the images, we identified six dynamic features, confirming their significant relevance to treatment response. These insights highlight the model's profound capacity for understanding and interpreting visual information and also showcase DAM's intricate structural design. In this context, the visual encoder adeptly focuses on extracting key image features, while the transformer-based feature aggregation efficiently manages the diverse distributions present within or among cellular images. This advanced method offers substantial potential for future explorations into deeper visual understanding patterns in AI models, providing invaluable assistance to physicians in interpreting more complex visual pattern information, thus enhancing diagnostic and therapeutic approaches in oncology. However, two limitations of DAM are notable: 1) While we have collected all the data currently available to the best of our ability, the limited dataset size requires further accumulation and multi-center validation; 2) While currently focused on longitudinal liquid biopsy data, the future expansion of DAM to include more dynamic data sources is crucial for its full potential exploration. 5. Conclusions In summary, this study represents a pioneering endeavor in utilizing AI technology to dynamically model longitudinal liquid biopsy data, marking a notable advancement towards precise response prediction in patients with GC. This research demonstrates the potential of AI in longitudinal liquid biopsies analysis for precision medicine and highlights its impressive ability to comprehend complex visual features, promising enhanced collaboration between AI and clinicians in clinical settings. Abbreviations Abbreviation Full form GC Gastric cancer AI Artificial intelligence CTC Circulating tumor cell CEC Circulating endothelial cell CT Computerized tomography TIM Temporal interaction module MLP Multilayer perceptron CNN Convolutional neural network DAM Dynamic-aware model ROC Receiver operating characteristic AUC Area under the curve AFP Alpha-fetoprotein CEA Carcinoembryonic antigen CA19-9 Carbohydrate antigen 19-9 CA72-4 Cancer antigen 72-4 CA125 Cancer antigen 125 NSE Neuron-specific enolase CR Complete response PR Partial response PD Progressive disease SD Stable disease RECIST Response evaluation criteria in solid tumors Declarations Ethics approval and consent to participate The Peking University Cancer Hospital Ethics Committee granted ethical approval for this study (approval number: 2020KT08). All participants or their legally authorized representatives provided informed consent. Competing interests All authors declare no competing interests. Authors’ contributions Zifan Chen, Yilin Li, Yang Chen, and Li Zhang designed the study; Yanyan Li, Yinlin Li, and Xujiao Feng collected the data. Zifan Chen, Jie Zhao, Huimin Liu, and Xinyu Nan analyzed the data; Zifan Chen, Jie Zhao, Bin Dong, and Li Zhang proposed the model; Yilin Li, Bin Dong, Lin Shen, Yang Chen, and Li Zhang supervised the work; Zifan Chen, Jie Zhao, Yanyan Li, Yilin Li, and Xinyu Nan wrote the original draft. Bin Dong, Lin Shen, Yang Chen, and Li Zhang revised the manuscript. Zifan Chen, Jie Zhao, and Yanyan Li contributed equally to this work. Zifan Chen, Jie Zhao, Yanyan Li, Yang Chen, and Li Zhang directly accessed and verified the underlying data reported in the manuscript. Acknowledgements The authors sincerely thank all the participants involved in this study for their contribution. Data availability The datasets will be available upon reasonable request emailed to the corresponding author ( [email protected] ). The code for the model development is available in Supplementary Materials. References Joshi, S. S. & Badgwell, B. D. Current treatment and recent progress in gastric cancer. CA Cancer J Clin 71 , 264-279 (2021). Miller, K. D. et al. Cancer treatment and survivorship statistics, 2022. CA Cancer J Clin 72 , 409-436 (2022). Bang, Y. J. et al. Adjuvant capecitabine and oxaliplatin for gastric cancer after D2 gastrectomy (CLASSIC): a phase 3 open-label, randomised controlled trial. Lancet 379 , 315-321 (2012). Al-Batran, S. E. et al. Perioperative chemotherapy with fluorouracil plus leucovorin, oxaliplatin, and docetaxel versus fluorouracil or capecitabine plus cisplatin and epirubicin for locally advanced, resectable gastric or gastro-oesophageal junction adenocarcinoma (FLOT4): a randomised, phase 2/3 trial. Lancet 393 , 1948-1957 (2019). Bang, Y. J. et al. Trastuzumab in combination with chemotherapy versus chemotherapy alone for treatment of HER2-positive advanced gastric or gastro-oesophageal junction cancer (ToGA): a phase 3, open-label, randomised controlled trial. Lancet 376 , 687-697 (2010). Janjigian, Y. Y. et al. Pembrolizumab plus trastuzumab and chemotherapy for HER2-positive gastric or gastro-oesophageal junction adenocarcinoma: interim analyses from the phase 3 KEYNOTE-811 randomised placebo-controlled trial. Lancet 402 , 2197-2208 (2023). Yu, Y. Y. et al. The pivotal application of patient-derived organoid biobanks for personalized treatment of gastrointestinal cancers. Biomark Res 10 , 73 (2022). Dagogo-Jack, I. & Shaw, A. T. Tumour heterogeneity and resistance to cancer therapies. Nat Rev Clin Oncol 15 , 81-94 (2018). Di Capua, D., Bracken-Clarke, D., Ronan, K., Baird, A. M. & Finn, S. The Liquid Biopsy for Lung Cancer: State of the Art, Limitations and Future Developments. Cancers (Basel) 13 (2021). Deng, Z., Wu, S., Wang, Y. & Shi, D. Circulating tumor cell isolation for cancer diagnosis and prognosis. EBioMedicine 83 , 104237 (2022). Coombes, R. C. et al. Personalized Detection of Circulating Tumor DNA Antedates Breast Cancer Metastatic Recurrence. Clin Cancer Res 25 , 4255-4263 (2019). Stergiopoulou, D. et al. Comprehensive liquid biopsy analysis as a tool for the early detection of minimal residual disease in breast cancer. Sci Rep 13 , 1258 (2023). Pei, H., Li, L., Han, Z., Wang, Y. & Tang, B. Recent advances in microfluidic technologies for circulating tumor cells: enrichment, single-cell analysis, and liquid biopsy for clinical applications. Lab Chip 20 , 3854-3875 (2020). Connal, S. et al. Liquid biopsies: the future of cancer early detection. J Transl Med 21 , 118 (2023). Edsjo, A. et al. Precision cancer medicine: Concepts, current practice, and future developments. J Intern Med 294 , 455-481 (2023). He, S. et al. New horizons in the identification of circulating tumor cells (CTCs): An emerging paradigm shift in cytosensors. Biosens Bioelectron 203 , 114043 (2022). Chiang, P. J. et al. Programmed Cell Death Ligand 1 Expression in Circulating Tumor Cells as a Predictor of Treatment Response in Patients with Urothelial Carcinoma. Biology (Basel) 10 (2021). Troncarelli Flores, B. C. et al. Molecular and Kinetic Analyses of Circulating Tumor Cells as Predictive Markers of Treatment Response in Locally Advanced Rectal Cancer Patients. Cells 8 (2019). He, K. et al. Decoding the glycoproteome: a new frontier for biomarker discovery in cancer. J Hematol Oncol 17 , 12 (2024). Lei, Y. et al. Combined detection of aneuploid circulating tumor-derived endothelial cells and circulating tumor cells may improve diagnosis of early stage non-small-cell lung cancer. Clin Transl Med 10 , e128 (2020). Ronzoni, M. et al. Circulating endothelial cells and endothelial progenitors as predictive markers of clinical response to bevacizumab-based first-line treatment in advanced colorectal cancer patients. Ann Oncol 21 , 2382-2389 (2010). Liu, Z. et al. Combination of tumor markers predicts progression and pathological response in patients with locally advanced gastric cancer after neoadjuvant chemotherapy treatment. BMC Gastroenterol 21 , 283 (2021). Machida, M. et al. Integrated Predictors by Propensity Scoring With Tumor Markers and Plasma Levels of microRNA-21-5p, IL-17, and IL-10 Complement Early Detection of Hepatocellular Carcinoma in Patients With Liver Cirrhosis. Technol Cancer Res Treat 22 , 15330338231212084 (2023). Salehi, M. et al. Circulating Tumor Cells as a Promising Tool for Early Detection of Hepatocellular Carcinoma. Cells 12 (2023). Wang, H. T. et al. Circulating tumor cells positivity provides an early detection of recurrence of pancreatic cancer. J Formos Med Assoc 122 , 653-655 (2023). Feng, Z. et al. Circulating tumor cells in the early detection of human cancers. Int J Biol Sci 18 , 3251-3265 (2022). Vanguri, R. S. et al. Multimodal integration of radiology, pathology and genomics for prediction of response to PD-(L)1 blockade in patients with non-small cell lung cancer. Nat Cancer 3 , 1151-1164 (2022). Boehm, K. M. et al. Multimodal data integration using machine learning improves risk stratification of high-grade serous ovarian cancer. Nat Cancer 3 , 723-733 (2022). Wang, Q., Li, Z., Zhai, W. & Zheng, J. Clinical values of circulating tumor cells count in localized renal cell carcinoma. Transl Cancer Res 12 , 2351-2360 (2023). Schoenberg, M. B. et al. Highly differential count of circulating and tumor infiltrating immune cells in patients with non-HCV/non-HBV hepatocellular carcinoma. Cancer Immunol Immunother 71 , 1103-1113 (2022). Syrigos, K., Fiste, O., Charpidou, A. & Grapsa, D. Circulating tumor cells count as a predictor of survival in lung cancer. Crit Rev Oncol Hematol 125 , 60-68 (2018). Devlin, J., Chang, M.-W., Lee, K. & Toutanova, K. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018). Brown, T. et al. Language models are few-shot learners. Advances in neural information processing systems 33 , 1877-1901 (2020). Vaswani, A. et al. Attention is all you need. Advances in neural information processing systems 30 (2017). He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition. 770-778. Dosovitskiy, A. et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020). Yuan, J. et al. Advanced prompting as a catalyst: Empowering large language models in the management of gastrointestinal cancers. The Innovation (2023). Acosta, J. N., Falcone, G. J., Rajpurkar, P. & Topol, E. J. Multimodal biomedical AI. Nat Med 28 , 1773-1784 (2022). Thirunavukarasu, A. J. et al. Large language models in medicine. Nat Med 29 , 1930-1940 (2023). Huang, Z., Bianchi, F., Yuksekgonul, M., Montine, T. J. & Zou, J. A visual-language foundation model for pathology image analysis using medical Twitter. Nat Med 29 , 2307-2316 (2023). Theodoris, C. V. et al. Transfer learning enables predictions in network biology. Nature 618 , 616-624 (2023). Rajpurkar, P., Chen, E., Banerjee, O. & Topol, E. J. AI in health and medicine. Nat Med 28 , 31-38 (2022). Johnson, K. B. et al. Precision Medicine, AI, and the Future of Personalized Health Care. Clin Transl Sci 14 , 86-93 (2021). He, J. et al. The practical implementation of artificial intelligence technologies in medicine. Nat Med 25 , 30-36 (2019). Song, Z. et al. Clinically applicable histopathological diagnosis system for gastric cancer detection using deep learning. Nat Commun 11 , 4294 (2020). He, M. et al. Associations of subcutaneous fat area and Systemic Immune-inflammation Index with survival in patients with advanced gastric cancer receiving dual PD-1 and HER2 blockade. J Immunother Cancer 11 (2023). He, M. et al. Deep learning model based on multi-lesion and time series CT images for predicting the benefits from anti-HER2 targeted therapy in stage IV gastric cancer. Insights Imaging 15 , 59 (2024). Li, J. et al. CT-based delta radiomics in predicting the prognosis of stage IV gastric cancer to immune checkpoint inhibitors. Front Oncol 12 , 1059874 (2022). Lu, L., Dercle, L., Zhao, B. & Schwartz, L. H. Deep learning for the prediction of early on-treatment response in metastatic colorectal cancer from serial medical imaging. Nat Commun 12 , 6654 (2021). Xu, Y. et al. Deep Learning Predicts Lung Cancer Treatment Response from Serial Medical Imaging. Clin Cancer Res 25 , 3266-3275 (2019). Li, Y. et al. Evolutionary expression of HER2 conferred by chromosome aneuploidy on circulating gastric cancer cells contributes to developing targeted and chemotherapeutic resistance. Clinical Cancer Research 24 , 5261-5271 (2018). Li, Y. et al. Dynamic monitoring of circulating tumour cells to evaluate therapeutic efficacy in advanced gastric cancer. British journal of cancer 114 , 138-145 (2016). Ng, A. Y. Preventing" overfitting" of cross-validation data. In: ICML. 245-253. Selvaraju, R. R. et al. Grad-cam: Visual explanations from deep networks via gradient-based localization. In: Proceedings of the IEEE international conference on computer vision. 618-626. Paszke, A. et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems 32 (2019). Supplementary Files 1Supplementaryinformation.docx 3Code.docx 4Data.xlsx Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-5181858","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":374858166,"identity":"b21302f8-3e18-4de4-8676-54c8519b0d33","order_by":0,"name":"zifan chen","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA7UlEQVRIiWNgGAWjYFACxoYPDAwScvzszQeAPAsGBh7CWhpnAFUaS/YcS2BgSJAgRgsDI1BLReKGGz4GxGnh7z/c2PBzhwRQC8/Hx7w/gC7kOcD44WMObi0SBw42NvaekTCeebt3szFPgoSxZG8Ds+TMbbi1GDA2tj/gbZOQ7btzdps0UEvihvMMbMy8+LQwMzY2/m2TYGy4kfOMSC1sjI3NQFsUJ9zIYYNoOduAX4vEGaAW2TYJUCAbG85JAzEONuP1C3//8YeNb9vqQFH58MEbGxtgiCUf/PARjxZsgLGBNPWjYBSMglEwCjAAAGXqUbANLYc1AAAAAElFTkSuQmCC","orcid":"","institution":"Peking University","correspondingAuthor":true,"prefix":"","firstName":"zifan","middleName":"","lastName":"chen","suffix":""},{"id":374858167,"identity":"0531298c-ae5c-4810-afed-1d31eb47a0e8","order_by":1,"name":"Jie Zhao","email":"","orcid":"","institution":"Peking University","correspondingAuthor":false,"prefix":"","firstName":"Jie","middleName":"","lastName":"Zhao","suffix":""},{"id":374858168,"identity":"9a19a931-ae3d-4444-ad41-e46a4d38c11c","order_by":2,"name":"Yanyan Li","email":"","orcid":"","institution":"Peking University Cancer Hospital: Beijing Cancer Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yanyan","middleName":"","lastName":"Li","suffix":""},{"id":374858169,"identity":"2b152e9c-698e-4ae2-a7dc-73fbebdc62eb","order_by":3,"name":"Yilin Li","email":"","orcid":"","institution":"Peking University Cancer Hospital: Beijing Cancer Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yilin","middleName":"","lastName":"Li","suffix":""},{"id":374858170,"identity":"a89be99b-5887-4cc5-9370-f0343f5c4c2d","order_by":4,"name":"Xinyu Nan","email":"","orcid":"","institution":"Peking University","correspondingAuthor":false,"prefix":"","firstName":"Xinyu","middleName":"","lastName":"Nan","suffix":""},{"id":374858171,"identity":"a324e3e2-ca80-4013-b01a-3ec8579eeb15","order_by":5,"name":"Huimin Liu","email":"","orcid":"","institution":"Guangzhou Medical University","correspondingAuthor":false,"prefix":"","firstName":"Huimin","middleName":"","lastName":"Liu","suffix":""},{"id":374858172,"identity":"d82dcf01-83ab-4ef9-becf-56bc4673eac7","order_by":6,"name":"Xujiao Feng","email":"","orcid":"","institution":"Peking University Cancer Hospital: Beijing Cancer Hospital","correspondingAuthor":false,"prefix":"","firstName":"Xujiao","middleName":"","lastName":"Feng","suffix":""},{"id":374858173,"identity":"6278d825-a2a2-4b81-845e-c6647b66d42f","order_by":7,"name":"Bin Dong","email":"","orcid":"","institution":"Peking University","correspondingAuthor":false,"prefix":"","firstName":"Bin","middleName":"","lastName":"Dong","suffix":""},{"id":374858174,"identity":"4b2b56ee-b124-415d-a24f-3937d720c62f","order_by":8,"name":"Lin Shen","email":"","orcid":"","institution":"Peking University Cancer Hospital: Beijing Cancer Hospital","correspondingAuthor":false,"prefix":"","firstName":"Lin","middleName":"","lastName":"Shen","suffix":""},{"id":374858175,"identity":"45c1a6a1-779f-435a-923f-a8efd0e8966c","order_by":9,"name":"Yang Chen","email":"","orcid":"https://orcid.org/0000-0001-6993-4870","institution":"Peking University Cancer Hospital: Beijing Cancer Hospital","correspondingAuthor":false,"prefix":"","firstName":"Yang","middleName":"","lastName":"Chen","suffix":""},{"id":374858176,"identity":"622677b6-3171-4dad-aaa3-af2a293ef151","order_by":10,"name":"Li Zhang","email":"","orcid":"","institution":"Peking University","correspondingAuthor":false,"prefix":"","firstName":"Li","middleName":"","lastName":"Zhang","suffix":""}],"badges":[],"createdAt":"2024-09-30 15:07:21","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5181858/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5181858/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":70477487,"identity":"e7912d01-e30b-4885-bd4d-66370609cf4d","added_by":"auto","created_at":"2024-12-03 14:26:39","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":247284,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eWorkflow to predict GC treatment response via dynamic-aware model (DAM) with longitudinal liquid biopsy data.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eA. \u003c/strong\u003eData format: Tumor-related aneuploid cellular images and indices for six tumor markers (AFP, CEA, CA199, CA72.4, CA125, and NSE) were collected from longitudinal liquid biopsies at various time points, forming dynamic patient profiles.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eB. \u003c/strong\u003eModel framework: DAM comprises five components: The cellular aggregator and tumor marker aggregator compile tumor-related cellular images and tumor marker indices into features, respectively. The temporal interaction module (TIM) integrates these mismatched multisource data. A temporal aggregator synthesizes this data into a patient-level feature for the predictor to classify patients as responders or non-responders.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eC.\u003c/strong\u003eExperimental pipeline: The study included 91 patients with GC from Peking Cancer Hospital, encompassing 1,895 tumor-related aneuploid cellular images and 1,698 tumor marker indices from longitudinal liquid biopsies. These patients were randomly divided into a training set of 74 patients and an independent test set of 17 patients. The training set was then subjected to three-fold (F1, F2, F3) cross-validation, resulting in three separate trained models (M1, M2, M3) and corresponding sets of validation results. Finally, these three trained models collaboratively predict responses on the independent test set, with their predictions being ensembled to yield the final result.\u003c/p\u003e","description":"","filename":"FIG1.png","url":"https://assets-eu.researchsquare.com/files/rs-5181858/v1/7c5b55d1cd26f7d782a814c3.png"},{"id":70477483,"identity":"4b64c1dd-6cb5-446d-917a-efd60166dd6c","added_by":"auto","created_at":"2024-12-03 14:26:39","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":151495,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eVisualization of longitudinal liquid biopsy data characteristics and dynamic distribution across patients with GC.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eA.\u003c/strong\u003e Bubble diagram shows the heterogeneity in tumor-related cellular images. Each column corresponds to an individual patient, while each row represents different time intervals, including Pre-Baseline (Pre-BL), within 45 days post-baseline ( ≤45d), within 3 months post-baseline ( ≤\u0026nbsp;3m), within 6 months post-baseline (≤ \u0026nbsp;6m), within a year post-baseline ( ≤\u0026nbsp;1y), and beyond a year post-baseline ( \u0026nbsp;\u0026gt;1y). The presence of each dot signifies the number of time points collected from a patient within these respective time intervals, with larger dots indicating a higher frequency of sampling. Bar graph adjacent to the right side of the panel further elucidates the frequency of sampling across these distinct time intervals.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eB.\u003c/strong\u003e Bubble diagram similarly depicts the heterogeneity in the distribution of tumor marker indices, employing an interpretation approach analogous to that used in \u003cstrong\u003eA\u003c/strong\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eC.\u003c/strong\u003e Box plot highlights the heterogeneity in cell counts across the aforementioned time intervals.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eD.\u003c/strong\u003e Box plot differentiates the heterogeneity in cell counts across time intervals between responders and non-responders.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eE.\u003c/strong\u003e Diagram illustrates the division of these time intervals.\u003c/p\u003e","description":"","filename":"FIG2.png","url":"https://assets-eu.researchsquare.com/files/rs-5181858/v1/1629cd7334e8c0a2afd5cc95.png"},{"id":70479382,"identity":"bb5b0d55-a1fe-4c78-93e3-5a431662eb78","added_by":"auto","created_at":"2024-12-03 14:42:39","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":115919,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eComparative performance analysis of GC treatment response prediction models.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eA.\u003c/strong\u003e Operating Characteristic Curves (ROC) of the cell-counting-based baseline model, utilizing cellular count data to predict treatment responses, across three-fold cross-validation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eB-C.\u003c/strong\u003e ROCs of two DAM-variants, DAM-TCI and DAM-TM, respectively.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eD.\u003c/strong\u003e ROC of the proposed DAM.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eE.\u003c/strong\u003e Robustness evaluation of DAM on an independent test set.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eF.\u003c/strong\u003e Time-series dynamic analysis of DAM from a span of forty-five days to all available data. The black lines on the bar chart depict the standard deviation.\u003c/p\u003e","description":"","filename":"FIG3.png","url":"https://assets-eu.researchsquare.com/files/rs-5181858/v1/b556452dc6344aa81bafdf90.png"},{"id":70477973,"identity":"682b54ee-c837-4193-9ae1-1ed5da28dae8","added_by":"auto","created_at":"2024-12-03 14:34:39","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":851605,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eVisual analysis of DAM’s attention mechanisms and focus areas.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eA. \u003c/strong\u003eBox plot showcases the distribution of important scores assigned to cellular features within the cellular aggregator, grouped by CTCs and CECs. The dark blue and orange color gradient indicates a progression from low to high significance respectively.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eB. \u003c/strong\u003eVisualization of focus areas by the model in tumor-related cellular images across time points, showing different quartile activations.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eC.\u003c/strong\u003e Box plots of six dynamic focus-area-related features, showing statistical distributions for responders and non-responders.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eD. \u003c/strong\u003eROCs of a dynamic-feature-driven nonlinear analysis via an MLP with three-fold cross-validation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eE. \u003c/strong\u003eRobustness evaluation of the nonlinear analysis on the independent test set.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eF.\u003c/strong\u003e Permutation importance analysis for six dynamic focus-area-related features.\u003c/p\u003e","description":"","filename":"FIG4.png","url":"https://assets-eu.researchsquare.com/files/rs-5181858/v1/4ae9db8c64cca1c40c2c53b8.png"},{"id":73321022,"identity":"c3d4793a-bb44-47f4-8339-18d42f9024b1","added_by":"auto","created_at":"2025-01-08 22:20:35","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2389876,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5181858/v1/6e61df03-f0ff-4feb-9f4e-5f98a958db48.pdf"},{"id":70477489,"identity":"f670ffe6-3efb-4530-934e-c9a4e357bc24","added_by":"auto","created_at":"2024-12-03 14:26:40","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":7510693,"visible":true,"origin":"","legend":"","description":"","filename":"1Supplementaryinformation.docx","url":"https://assets-eu.researchsquare.com/files/rs-5181858/v1/2cc042831843057b07a04a10.docx"},{"id":70477972,"identity":"fb121a93-164d-455b-85ec-203c87c69d5c","added_by":"auto","created_at":"2024-12-03 14:34:39","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":13242,"visible":true,"origin":"","legend":"","description":"","filename":"3Code.docx","url":"https://assets-eu.researchsquare.com/files/rs-5181858/v1/4e48017d34b0569162afecac.docx"},{"id":70477484,"identity":"c8f4d599-66c3-4c41-ab20-5a79b115cd73","added_by":"auto","created_at":"2024-12-03 14:26:39","extension":"xlsx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":13510,"visible":true,"origin":"","legend":"","description":"","filename":"4Data.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-5181858/v1/860ff17cd6672bc726b06a3e.xlsx"}],"financialInterests":"","formattedTitle":"Predicting response to patients with gastric cancer via dynamic-aware model with longitudinal liquid biopsy data","fulltext":[{"header":"1. Background","content":"\u003cp\u003eGastric cancer (GC), ranking fifth globally in prevalence, stands as the third leading cause of cancer-related mortality\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e,\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. The widespread occurrence of GC demands the formulation of effective treatments and precise predictive models for treatment responses\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. However, GC treatment is challenged by patient heterogeneity, hindering accurate treatment response predictions\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e,\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. Such heterogeneity, rooted in cellular characteristics and morphological variations, is critical in devising patient-specific treatment plans\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e,\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. In response, recent advancements in analyzing liquid biopsy data, including circulating tumor cells (CTCs), circulating endothelial cells (CECs), and tumor markers, have emerged as a potent tool in cancer treatment management\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e,\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. Analysis of tumor-related cells and tumor markers in liquid biopsies offers dual advantages: firstly, it acts as a cellular-level biomarker, providing insights into cellular heterogeneity and its association with treatment response\u003csup\u003e\u003cspan additionalcitationids=\"CR12 CR13\" citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e; secondly, liquid biopsies are more convenient than traditional diagnostics like computerized tomography (CT) scans and histopathological examinations\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e,\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e, facilitating the gathering of time-sensitive data essential for understanding a patient\u0026rsquo;s evolving cellular changes during treatment\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. Indeed, longitudinal liquid biopsy data is expected to provide personalized treatment strategies to adapt to the changing nature of the disease and patient response.\u003c/p\u003e \u003cp\u003eRecent advancements in cancer diagnosis and treatment management have largely benefited from cellular biomarker applications, including CTCs\u003csup\u003e\u003cspan additionalcitationids=\"CR17 CR18\" citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e, CECs\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e,\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e, and tumor markers\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e,\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e. These developments have substantially improved our understanding of cancer, from early detection\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e,\u003cspan additionalcitationids=\"CR25\" citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e to treatment outcome prediction\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e,\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e,\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e,\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e,\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e. However, many previous studies\u003csup\u003e\u003cspan additionalcitationids=\"CR30\" citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e have relied on cell-counting-based statistical approaches for outcome assessment. While useful, these methods frequently overlook potential insights that longitudinal liquid biopsy data can provide. Conversely, artificial intelligence (AI), particularly in sequence modeling\u003csup\u003e\u003cspan additionalcitationids=\"CR33\" citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e and visual information extraction\u003csup\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e,\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e, shows promise in revolutionizing biomedical data analysis\u003csup\u003e\u003cspan additionalcitationids=\"CR38 CR39 CR40\" citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e\u003c/sup\u003e. AI's ability to process large datasets and decode complex patterns may lead to more precise and personalized treatment efficacy assessments\u003csup\u003e\u003cspan additionalcitationids=\"CR43\" citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e\u003c/sup\u003e. Integrating AI into GC management enables the creation of uniquely tailored treatment strategies for each patient's molecular profile, enhancing treatment efficacy and minimizing invasiveness\u003csup\u003e\u003cspan additionalcitationids=\"CR46\" citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e\u003c/sup\u003e. Although some studies\u003csup\u003e\u003cspan additionalcitationids=\"CR49\" citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e\u003c/sup\u003e have begun exploring the role of dynamic data in cancer treatment management, to our knowledge, none have yet investigated longitudinal liquid biopsies, especially regarding dynamic tumor-related cellular images and dynamic tumor markers. This highlights the need for more advanced analytical methods in longitudinal liquid biopsy data analysis.\u003c/p\u003e \u003cp\u003eAddressing these challenges, we compiled a comprehensive longitudinal dataset from 91 patients with GC treated at Peking Cancer Hospital between July 2019 and April 2022. This dataset includes 1,895 aneuploid tumor-related cellular images and 1,698 tumor marker indices from six markers, derived from longitudinal liquid biopsies collected across all available follow-up visits. We randomly divided the dataset into a training set consisting of 74 patients, using three-fold cross-validation for model development, and a test set of 17 patients to assess the model\u0026rsquo;s generalizability and robustness. We developed a deep-learning-based dynamic-aware model (DAM) to precisely predict GC treatment responses. DAM uniquely tackles the challenge of interpreting complex patterns and temporal dynamics in treatment response predictions by combining convolutional and fully connected neural networks for feature extraction with attention mechanisms for information integration. Specifically, it uses self-attention-based modules to integrate multi-object and multi-temporal data, and a cross-attention-based module to merge mismatched multisource dynamic data. Experimental results confirm DAM\u0026rsquo;s efficacy in deriving insights from longitudinal liquid biopsies and accurately predicting treatment responses. Furthermore, we identified six dynamic focus area features via DAM\u0026rsquo;s visual analysis and conducted preliminary studies to assess its potential for interpreting liquid biopsy visual data.\u003c/p\u003e"},{"header":"2. Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1. Ethics\u003c/h2\u003e \u003cp\u003e The Peking University Cancer Hospital Ethics Committee granted ethical approval for this study (approval number: 2020KT08). All participants or their legally authorized representatives provided informed consent.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2. Patients and data collection\u003c/h2\u003e \u003cp\u003eThe study included patients with GC from July 2019 to April 2022. For each patient, longitudinal liquid biopsies were collected, comprising dynamic tumor-related aneuploid cellular images (CTCs and CECs) and dynamic tumor marker indices, at baseline and during subsequent follow-ups (Figs.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eA and S1). Blood samples underwent density gradient centrifugation and microfluidic isolation to enrich CTCs and CECs. Subsequently, isolated cells were fixed onto slides, and stained with specific markers, including cluster of differentiation 31 (CD31), cluster of differentiation 45 (CD45), centromere protein 8 (CEP8), and 4',6-diamidino-2-phenylindole (DAPI), using specific iFISH (immunostaining-FISH) staining following the manufacturer\u0026rsquo;s protocol\u003csup\u003e\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e\u003c/sup\u003e (Cytelligen, San Diego, CA, USA) with minor alteration, and then imaged using the automated Metafer-i\u0026bull;FISH\u0026reg; CTC 3D scanning and image analysis system\u003csup\u003e\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e\u003c/sup\u003e codeveloped by Carl Zeiss (Oberkochen, Germany), MetaSystems (Altlussheim, Germany) and Cytelligen to capture high-resolution multi-channel overlay images for analysis (More details in Text S1). Furthermore, blood samples were analyzed in the laboratory to measure levels of various tumor markers, including alpha-fetoprotein (AFP), carcinoembryonic antigen (CEA), carbohydrate antigen 19\u0026thinsp;\u0026minus;\u0026thinsp;9 (CA19-9), cancer antigen 72\u0026thinsp;\u0026minus;\u0026thinsp;4 (CA72-4), cancer antigen (CA125), and neuron-specific enolase (NSE), constituting the collected tumor marker indices.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn patients with GC, the definition of treatment response is as follows: Patients who achieve a complete response (CR) or partial response (PR) according to the Response Evaluation Criteria in Solid Tumors (RECIST) are categorized as responders. Conversely, those exhibiting progressive disease (PD) or stable disease (SD) are classified as non-responders.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3. The architecture of DAM\u003c/h2\u003e \u003cp\u003eWe introduced the DAM, a deep-learning-based dynamic-aware model designed to analyze dynamic tumor-related cellular images and dynamic tumor marker indices from longitudinal liquid biopsies to predict GC treatment responses. DAM\u0026rsquo;s architecture, shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eB, comprises a comprehensive system with five integrated components. DAM begins with the cellular aggregator, employing a ResNet-18\u003csup\u003e35\u003c/sup\u003e, as the feature extractor and a dual-stage Transformer\u003csup\u003e\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e, to create a unified representation of cellular data across various time points (Fig. \u003cspan refid=\"MOESM2\" class=\"InternalRef\"\u003eS2\u003c/span\u003e). Concurrently, the tumor marker aggregator utilizes a dual-layer perceptron and a similar dual-stage Transformer for the integration of tumor marker indices over time (Fig. \u003cspan refid=\"MOESM3\" class=\"InternalRef\"\u003eS3\u003c/span\u003e). Subsequently, the temporal interaction module (TIM) aligns these dynamic features using an advanced cross-attention mechanism (Fig. \u003cspan refid=\"MOESM4\" class=\"InternalRef\"\u003eS4\u003c/span\u003e), ensuring effective synchronization and integration of mismatched multisource temporal data. Following this, the temporal aggregator employs a quad-stage Transformer to consolidate data from various time points into a comprehensive patient-centric feature representation (Fig. S5). Lastly, the predictor component utilizes a three-layer multilayer perceptron (MLP) to classify patients into responder or non-responder categories based on the integrated features (Fig. S6). All of these components operate in concert to ensure a robust and accurate analysis of longitudinal liquid biopsies.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4. Training and assessment procedures of DAM\u003c/h2\u003e \u003cp\u003eThis study meticulously designed experiments to ensure DAM\u0026rsquo;s effective training and fair verification using a constrained dataset (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eC). The dataset was randomly divided, allocating 74 patients for training and a separate set of 17 patients reserved for independent testing. The training dataset was further subdivided into three subsets: fold-1 (F1), fold-2 (F2), and fold-3 (F3), comprising 26, 26, and 22 patients with GC, respectively. A three-fold cross-validation strategy\u003csup\u003e\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e\u003c/sup\u003e was implemented, rotating these subsets between training and validation to fine-tune the model's hyperparameters and architecture. This process yielded three distinct trained models, collectively forming an ensemble model evaluated against the independent test set for robustness.\u003c/p\u003e \u003cp\u003eTo enhance the model's applicability in real-world medical scenarios, we introduced a dynamic longitudinal elimination strategy during training (Fig. S7). This strategy entailed an 80% non-repetitive random sampling along the temporal dimension of the dynamic tumor-related cellular images and dynamic tumor markers, effectively mimicking a data-level dropout. This strategy fosters model resilience and adaptability to practical challenges like missing some temporal data. When validation and testing, we utilized all available longitudinal data without any random sampling to ensure a thorough assessment.\u003c/p\u003e \u003cp\u003eAfter training, we froze model parameters for further analysis. We first applied softmax activation function to predictive probability of DAM and used them to plot receiver operating characteristic (ROC) curves and calculate an area under the curve (AUC) scores. Furthermore, employing the GradCAM algorithm \u003csup\u003e\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e\u003c/sup\u003e on the final convolutional layer of the cellular aggregator produced insightful attention maps. Based on these maps, six related dynamic features were quantified (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e4\u003c/span\u003eA). Specifically, for a particular patient with a total of \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:N\\)\u003c/span\u003e\u003c/span\u003e tumor-related cellular images \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\{{I}_{1},{I}_{2},\\cdots\\:,{I}_{N}\\}\\)\u003c/span\u003e\u003c/span\u003e across \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:T\\:\\)\u003c/span\u003e\u003c/span\u003etime points, each image is associated with a corresponding attention map generated by GradCAM, denoted as \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\{{A}_{1},{A}_{2},\\cdots\\:,{A}_{N}\\}\\)\u003c/span\u003e\u003c/span\u003e. These attention maps were then subjected to binary segmentation using a 75-th percentile threshold to delineate focus areas \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\{{F}_{1},{F}_{2},\\cdots\\:,{F}_{N}\\}\\)\u003c/span\u003e\u003c/span\u003e, where the focus area is defined as the total area identified as foreground in the binary segmentation. The dynamic features, namely, focus area variability (VarFA), minimum focus area (MinFA), maximum focus area (MaxFA), average focus area (AvgFA), and median focus area (MedFA), were denoted as follows,\u003c/p\u003e \u003cp\u003e \u003cdiv id=\"Equa\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:\\text{A}\\text{v}\\text{g}\\text{F}\\text{A}=\\frac{1}{N}{\\sum\\:}_{i=1}^{N}{F}_{i}^{\\:},$$\u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Equb\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:\\text{V}\\text{a}\\text{r}\\text{F}\\text{A}=\\frac{1}{N-1}{\\sum\\:}_{i=1}^{N}{\\left({F}_{i}-AvgFA\\right)}^{2},$$\u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Equc\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\:\\text{M}\\text{i}\\text{n}\\text{F}\\text{A}=\\text{m}\\text{i}\\text{n}\\{{F}_{1},{F}_{2},\\cdots\\:,{F}_{N}\\},$$\u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Equd\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$$\\:\\text{M}\\text{a}\\text{x}\\text{F}\\text{A}=\\text{m}\\text{a}\\text{x}\\{{F}_{1},{F}_{2},\\cdots\\:,{F}_{N}\\},$$\u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Eque\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Eque\" name=\"EquationSource\"\u003e\n$$\\:\\text{M}\\text{e}\\text{d}\\text{F}\\text{A}=\\text{m}\\text{e}\\text{d}\\text{i}\\text{a}\\text{n}\\left\\{{F}_{1},{F}_{2},\\cdots\\:,{F}_{N}\\right\\}.$$\u003c/div\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003eAdditionally, focus area dispersion (DisFA) was denoted by determining the number of connected domains within the regions of interest, expressed as\u003cdiv id=\"Equf\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equf\" name=\"EquationSource\"\u003e\n$$\\:\\text{D}\\text{i}\\text{s}\\text{F}\\text{A}={\\sum\\:}_{i=1}^{N}{D}_{i},$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{D}_{i}\\)\u003c/span\u003e\u003c/span\u003e represents the number of connected domains in the \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:i\\)\u003c/span\u003e\u003c/span\u003e-th image. Finally, a three-layer MLP, was trained using these features to uncover nonlinear relationships between the features and treatment responses.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5. Statistical analyses\u003c/h2\u003e \u003cp\u003eThe sample size determination was contingent upon the number of patients fulfilling the inclusion criteria (first- or second-line treatment with at least two time-point data) rather than a pre-established statistical methodology. The allocation of subjects across different groups, as delineated in Figs.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e2\u003c/span\u003eC\u0026ndash;D, \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e4\u003c/span\u003eA, and \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e4\u003c/span\u003eC, was evaluated utilizing the Mann-Whitney U test. We conducted our statistical analyses via R software (version 4.1.3) or Python (version 3.7.10). A \u003cem\u003eP\u003c/em\u003e-value threshold below 0.05 was designated as the criterion for statistical significance.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo ensure the reproducibility of this study, comprehensive methodologies and data management protocols are thoroughly documented in the supplementary information (Figs. S1\u0026ndash;S10 and Texts S1\u0026ndash;S3). The robustness of DAM was augmented through a three-fold cross-validation strategy and was further validated by an independent test set. Moreover, the source code, implemented by PyTorch\u003csup\u003e\u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e\u003c/sup\u003e, is available in the supplementary materials.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.6. Role of funders\u003c/h2\u003e \u003cp\u003eThe funders played no role in the study design, data collection, data analysis, data interpretation, and writing of the report.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.1. Patient profile and data characteristics\u003c/h2\u003e \u003cp\u003eThe data in this study were collected from patients with GC treated at Peking Cancer Hospital during the period between July 2019 and April 2022. The dataset comprises longitudinal liquid biopsy data from 91 patients, encompassing 1,895 aneuploid tumor-related cellular images and 1,698 tumor marker indices. The dataset was randomly divided into two a training set of 74 patients for deep-learning model development, using three-fold cross-validation, and a test set of the remaining 17 patients to assess the model's generalizability and robustness (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). The median age of the patients was 65 years, with an interquartile range of 57 to 72 years. Males constituted 80.22% of the dataset, indicating a pronounced male predominance. Treatment lines included first-line therapies (80.22%) and other-line therapies (19.78%). Most patients were diagnosed with advanced-stage disease.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePathological characteristics of enrolled patients.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003ePatient baseline characteristics\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colspan=\"5\" nameend=\"c7\" namest=\"c3\"\u003e \u003cp\u003e\u003cem\u003eN\u003c/em\u003e (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e\u003cem\u003eP\u003c/em\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eFold-1\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eFold-2\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eFold-3\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eTest\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAge\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026ge;60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e12 (46.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e17 (65.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e12 (54.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e11 (64.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.482\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u0026lt;60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e14 (53.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e9 (34.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e10 (45.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e6 (35.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSex\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e22 (84.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e20 (76.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e20 (90.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e11 (64.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.199\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFemale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e4 (15.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6 (23.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e2 (9.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e6 (35.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eLocation\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGEJ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e9 (34.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e8 (30.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e7 (31.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e5 (29.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.985\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNon-GEJ\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e17 (65.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e18 (69.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e15 (68.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e12 (70.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eDifferentiation\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHigh\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1 (3.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1 (4.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"5\" rowspan=\"6\"\u003e \u003cp\u003e0.562\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModerate-high\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1 (4.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModerate\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e11 (42.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e13 (50.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e6 (27.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e6 (35.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eModerate-poor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e7 (26.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5 (19.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e6 (27.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2 (11.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePoor\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e8 (30.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e7 (26.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e7 (31.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e9 (52.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUnknown\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1 (4.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eLauren classification\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIntestinal type\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e17 (65.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e17 (65.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e14 (63.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e13 (76.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003e0.818\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiffused type\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e4 (15.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2 (7.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e3 (13.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e3 (17.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMixed type\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e4 (15.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4 (15.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e4 (18.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1 (5.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eUnknown\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1 (3.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3 (11.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1 (4.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eHER2 expression (IHC)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePositive\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e17 (65.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e18 (69.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e17 (77.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e13 (76.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003e0.775\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e9 (34.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e8 (30.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e5 (22.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e4 (23.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTherapeutic efficacy\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1 (3.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0 (0.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\" morerows=\"3\" rowspan=\"4\"\u003e \u003cp\u003e0.568\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePR\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e17 (65.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e13 (50.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e15 (68.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e8 (47.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e4 (15.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e7 (26.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e6 (27.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e7 (41,2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePD\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e5 (19.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5 (19.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e1 (4.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e2 (11.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"8\"\u003eNote: *\u003cem\u003eAbbreviations\u003c/em\u003e: GEJ: Esophagogastric junction; IHC: Immunohistochemical; CR: Complete response; PR: Partial response; SD: Stable disease; PD: Progressive disease.\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"8\"\u003e\u003cb\u003eA.\u003c/b\u003e Bubble diagram shows the heterogeneity in tumor-related cellular images. Each column corresponds to an individual patient, while each row represents different time intervals, including Pre-Baseline (Pre-BL), within 45 days post-baseline (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\le\\:\\)\u003c/span\u003e\u003c/span\u003e45d), within 3 months post-baseline (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\le\\:\\)\u003c/span\u003e\u003c/span\u003e3m), within 6 months post-baseline (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\le\\:\\)\u003c/span\u003e\u003c/span\u003e6m), within a year post-baseline (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\le\\:\\)\u003c/span\u003e\u003c/span\u003e1y), and beyond a year post-baseline (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\u0026gt;\\)\u003c/span\u003e\u003c/span\u003e1y). The presence of each dot signifies the number of time points collected from a patient within these respective time intervals, with larger dots indicating a higher frequency of sampling. Bar graph adjacent to the right side of the panel further elucidates the frequency of sampling across these distinct time intervals.\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"8\"\u003e\u003cb\u003eB.\u003c/b\u003e Bubble diagram similarly depicts the heterogeneity in the distribution of tumor marker indices, employing an interpretation approach analogous to that used in \u003cb\u003eA\u003c/b\u003e.\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"8\"\u003e\u003cb\u003eC.\u003c/b\u003e Box plot highlights the heterogeneity in cell counts across the aforementioned time intervals.\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"8\"\u003e\u003cb\u003eD.\u003c/b\u003e Box plot differentiates the heterogeneity in cell counts across time intervals between responders and non-responders.\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"8\"\u003e\u003cb\u003eE.\u003c/b\u003e Diagram illustrates the division of these time intervals.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe study meticulously tracked the temporal distribution of liquid biopsy metrics to capture the dynamic aspects of GC progression. Figures\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e2\u003c/span\u003eA\u0026ndash;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e2\u003c/span\u003eB show the dynamic distribution of tumor-related cellular images and tumor marker indices across patients and time intervals, respectively. The median number of collection time points per patient was three for both tumor-related cellular images and tumor marker indices, with an interquartile range of 1 to 4. This median indicates robust longitudinal data collection, providing a solid foundation for disease progression analysis. Additionally, Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e2\u003c/span\u003eC highlights fluctuations in tumor-related cell counts over different time intervals (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e2\u003c/span\u003eE), revealing consistent distribution patterns across all intervals. This consistency, observed regardless of treatment response (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e2\u003c/span\u003eD), underscores the nuanced complexity of disease trajectories reflected in liquid biopsy profiles.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.2. Enhanced response prediction using the dynamic-aware model compared to the cell-counting-based model\u003c/h2\u003e \u003cp\u003eComprehensive analysis shows that the proposed DAM outperforms the traditional cell-counting-based model in predicting GC treatment responses (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e3\u003c/span\u003e). The ROCs from three-fold cross-validation shows that fold-1 achieved an AUC of 0.739 (95% Confidence Interval [CI]: 0.528\u0026ndash;0.922), fold-2 an AUC of 0.845 (95% CI: 0.655\u0026ndash;0.976), and fold-3 an AUC of 0.838 (95% CI: 0.639\u0026ndash;0.979) (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e3\u003c/span\u003eD). The overall average AUC inclusive of the standard deviation was 0.807\u0026thinsp;\u0026plusmn;\u0026thinsp;0.048. Furthermore, applied to an independent test set, DAM demonstrated robust performance with an AUC of 0.802 (95% CI: 0.532\u0026ndash;1.000; Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e3\u003c/span\u003eE). By comparison, the cell-counting-based model, using the XGBoost algorithm based on baseline tumor-related cell counts recorded an average AUC of 0.582\u0026thinsp;\u0026plusmn;\u0026thinsp;0.037 (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e3\u003c/span\u003eA and Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo examine the impact of two longitudinal data sources, including dynamic tumor-related cellular images and dynamic tumor markers, we developed two DAM variants: DAM-TCI and DAM-TM. DAM-TCI focuses on solely utilizing dynamic tumor-related cellular images to predict response; while DAM-TM focuses on solely utilizing dynamic tumor markers. In our experiments, DAM-TCI achieved an average AUC of 0.731\u0026thinsp;\u0026plusmn;\u0026thinsp;0.047 (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e3\u003c/span\u003eB), whereas DAM-TM achieved an average AUC of 0.725\u0026thinsp;\u0026plusmn;\u0026thinsp;0.058 (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e3\u003c/span\u003eC). Both variants outperformed the cell-counting-based model, highlighting DAM\u0026rsquo;s dynamic design for synthesizing dynamic data from different time points.\u003c/p\u003e \u003cp\u003eThe mismatched collection times for dynamic tumor-related cellular images and tumor marker indices (Fig. \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e) necessitate the model\u0026rsquo;s ability to adaptively integrate mismatched multisource dynamic data. We proposed a enhanced cross-attention module (TIM) to deal with this challenge. Integrating dynamic tumor-related cellular images with dynamic tumor marker indices via TIM enhanced performance to 0.807 (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e3\u003c/span\u003eD), surpassing the individual dynamic data sources. This outcome underscores the efficacy of TIM in amalgamating multiple dynamic data types and the complementary role of tumor marker information in tumor-related cellular images.\u003c/p\u003e \u003cp\u003eSubsequently, we performed a time-series dynamic analysis of DAM, allowing the model to access data from specified time intervals, including forty-five days (45d), three months (3m), six months (6m), one year (1y), and all available data (all) in the collection (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e2\u003c/span\u003eE). DAM\u0026rsquo;s performance gradually improved with extended time intervals, showing average AUCs of gradual improvement in DAM\u0026rsquo;s performance with the extension of the time interval (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e3\u003c/span\u003eF), showing average AUCs of 0.644 (45d), 0.718 (3m), 0.786 (6m), 0.786 (1y), and 0.807 (all). This indicates that including more dynamic information from extended treatment records improves prediction accuracy. Notably, DAM showed modest improvement after six months (an absolute increase of 0.021, from 0.786 to 0.807), in contrast to a big improvement within the first six months (an absolute increase of 0.142, from 0.644 to 0.786). This underscores DAM\u0026rsquo;s effectiveness in using early treatment data for precise treatment response prediction, aligning with the clinical need to identify non-responders promptly in the treatment process.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.3. Visual analysis of DAM\u003c/h2\u003e \u003cp\u003eThe previously cell-counting-based approach overlooks the visual characteristics of tumor-related cellular images. In contrast, the proposed DAM enhances complex visual data interpretation, enabling a more comprehensive image analysis. We explored DAM\u0026rsquo;s visual focus mechanism by analyzing important score distributions for cellular features within the aggregator, categorized by CTCs and CECs (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e4\u003c/span\u003eA). CTCs had significantly higher important scores than CECs (two-sided Mann-Whitney U test, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.0001), emphasizing the crucial role of CTCs in DAM. Consequently, we visualized tumor-related cellular images at the 25-th, 50-th, and 75-th percentile of the important score distribution. Our intuitive observations suggest that larger and more semantically complex CTCs have a more substantial impact on outcomes. This leads us to believe that DAM primarily interprets liquid biopsy visual information through two key visual signals: size and heterogeneity within the image.\u003c/p\u003e \u003cp\u003eFurthermore, utilizing the GradCAM algorithm, we visualized DAM\u0026rsquo;s visual encoder, noting variations in focus area sizes and dispersion across different cellular images (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e4\u003c/span\u003eB). Building on DAM\u0026rsquo;s visual focus analysis insights, we further explored dynamic focus-area-related visual features (more details in the \u003cspan refid=\"Sec2\" class=\"InternalRef\"\u003emethods\u003c/span\u003e section), including VarFA, MinFA, MaxFA, AvgFA, MedFA, and DisFA (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e4\u003c/span\u003eC). We employed a nonlinear analysis via a three-layer MLP to predict treatment response using these dynamic visual features (Fig. S6), yielding AUC scores of 0.712 (95% CI: 0.444\u0026ndash;0.948) in fold-1, 0.804 (95% CI: 0.607\u0026ndash;0.955) in fold-2, and 0.819 (95% CI: 0.614\u0026ndash;0.983) in fold-3, with an average of 0.778\u0026thinsp;\u0026plusmn;\u0026thinsp;0.047 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e4\u003c/span\u003eC), with a comparable performance to DAM on the test set (0.806 vs. 0.802; Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e4\u003c/span\u003eD). Moreover, a permutation importance analysis highlighted AvgFA as particularly important, underscoring tumor cell size\u0026rsquo;s importance in DAM. These analyses offer initial insights into the visual characteristics of tumor-related cellular images, suggesting deep-learning models like DAM have the potential to summarize generalized spatial visual information.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eGC remains a significant global health challenge, underscoring the need for precise and early evaluation of treatment effectiveness. Recent studies on liquid biopsy data, including CTCS, CECs, and tumor markers, have highlighted its potential in cancer treatment management. Analyzing liquid biopsy data offers two main advantages: providing crucial cellular-level biomarkers and being more convenient than traditional methods like CT scans and histopathology. Such convenience aids in the acquisition of time-sensitive data. However, a notable research gap exists in processing and interpreting complex, longitudinal liquid biopsy data, highlighting the need for more advanced analytical methods. Our study is the first to apply longitudinal liquid biopsy data in developing an innovative AI model, DAM. DAM utilizes cutting-edge AI technology for integrating multi-source, multi-object, and multi-temporal data and makes accurate response predictions, signifying a considerable advancement in the field.\u003c/p\u003e \u003cp\u003eA key finding, distinguishing our research from previous liquid biopsy studies, is DAM\u0026rsquo;s superior performance in modeling longitudinal liquid biopsies for treatment response predictions. Unlike previously cell-counting-based methods, DAM processes dynamic data from CTCs, CECs, tumor markers, and associated visual imagery concurrently. The comparative analysis underscores DAM\u0026rsquo;s enhanced capability in prediction response in GC treatment, and an evaluation in the independent test set reveals its robustness. Additionally, two DAM variants, DAM-TCI and DAM-TM, underscore the importance of longitudinal information and demonstrate dynamic modeling\u0026rsquo;s effectiveness in DAM. The analysis of the impact of dynamic data from diverse treatment time intervals on efficacy prediction reveals that DAM can accurately predict outcomes using early treatment data, aligning closely with clinical usage scenarios.\u003c/p\u003e \u003cp\u003eThis study also showcases DAM\u0026rsquo;s module design\u0026rsquo;s exceptional flexibility and expansiveness, adeptly aggregating diverse elements like tumor-related cellular images. Additionally, the TIM within DAM demonstrates exceptional proficiency in aligning temporally mismatched data from diverse dynamic data sources. By utilizing tumor markers to enhance the analysis of tumor-related cellular images in this study, TIM establishes a groundbreaking approach for data integration. This method is particularly suited for medical contexts, where temporal discrepancies between datasets are common. Its applicability to future scenarios suggests a broad utility beyond the immediate context, offering a promising avenue for integrating diverse data types across a wide range of fields. Besides, within DAM, the temporal aggregator skillfully handles a varied temporal distribution of dynamic data. This innovative design not only accommodates multiple objects, times, and modalities but also seamlessly incorporates non-dynamic (single-time-point) data, such as historical radiological and pathological records. The versatility and scalability inherent in the module design are crucial, facilitating the adaptation and integration of a wide spectrum of data types and sources, thereby enhancing the model's applicability and robustness in complex clinical scenarios.\u003c/p\u003e \u003cp\u003eBy leveraging advanced deep learning technologies like convolutional neural networks (CNNs) and Transformers, DAM effectively models visual information of tumor-related cellular images, a facet overlooked by previously cell-counting-based methods. This approach facilitated a preliminary analysis of DAM's visual attention patterns. Through detailed evaluations of areas of interest of DAM within the images, we identified six dynamic features, confirming their significant relevance to treatment response. These insights highlight the model's profound capacity for understanding and interpreting visual information and also showcase DAM's intricate structural design. In this context, the visual encoder adeptly focuses on extracting key image features, while the transformer-based feature aggregation efficiently manages the diverse distributions present within or among cellular images. This advanced method offers substantial potential for future explorations into deeper visual understanding patterns in AI models, providing invaluable assistance to physicians in interpreting more complex visual pattern information, thus enhancing diagnostic and therapeutic approaches in oncology.\u003c/p\u003e \u003cp\u003eHowever, two limitations of DAM are notable: 1) While we have collected all the data currently available to the best of our ability, the limited dataset size requires further accumulation and multi-center validation; 2) While currently focused on longitudinal liquid biopsy data, the future expansion of DAM to include more dynamic data sources is crucial for its full potential exploration.\u003c/p\u003e"},{"header":"5. Conclusions","content":"\u003cp\u003eIn summary, this study represents a pioneering endeavor in utilizing AI technology to dynamically model longitudinal liquid biopsy data, marking a notable advancement towards precise response prediction in patients with GC. This research demonstrates the potential of AI in longitudinal liquid biopsies analysis for precision medicine and highlights its impressive ability to comprehend complex visual features, promising enhanced collaboration between AI and clinicians in clinical settings.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eAbbreviation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eFull form\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eGC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eGastric cancer\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eAI\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eArtificial intelligence\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eCTC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eCirculating tumor cell\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eCEC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eCirculating endothelial cell\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eCT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eComputerized tomography\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eTIM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eTemporal interaction module\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eMLP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eMultilayer perceptron\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eCNN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eConvolutional neural network\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eDAM\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eDynamic-aware model\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eROC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eReceiver operating characteristic\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eAUC\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eArea under the curve\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eAFP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eAlpha-fetoprotein\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eCEA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eCarcinoembryonic antigen\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eCA19-9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eCarbohydrate antigen 19-9\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eCA72-4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eCancer antigen 72-4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eCA125\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eCancer antigen 125\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eNSE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eNeuron-specific enolase\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eCR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eComplete response\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003ePR\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003ePartial response\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003ePD\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eProgressive disease\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eSD\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eStable disease\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 24.5098%;\"\u003e\n \u003cp\u003eRECIST\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 75.4902%;\"\u003e\n \u003cp\u003eResponse evaluation criteria in solid tumors\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eEthics approval and consent to participate\u003c/h2\u003e \u003cp\u003eThe Peking University Cancer Hospital Ethics Committee granted ethical approval for this study (approval number: 2020KT08). All participants or their legally authorized representatives provided informed consent.\u003c/p\u003e \u003c/p\u003e\u003cp\u003e \u003ch2\u003eCompeting interests\u003c/h2\u003e \u003cp\u003eAll authors declare no competing interests.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eAuthors\u0026rsquo; contributions\u003c/h2\u003e \u003cp\u003eZifan Chen, Yilin Li, Yang Chen, and Li Zhang designed the study; Yanyan Li, Yinlin Li, and Xujiao Feng collected the data. Zifan Chen, Jie Zhao, Huimin Liu, and Xinyu Nan analyzed the data; Zifan Chen, Jie Zhao, Bin Dong, and Li Zhang proposed the model; Yilin Li, Bin Dong, Lin Shen, Yang Chen, and Li Zhang supervised the work; Zifan Chen, Jie Zhao, Yanyan Li, Yilin Li, and Xinyu Nan wrote the original draft. Bin Dong, Lin Shen, Yang Chen, and Li Zhang revised the manuscript. Zifan Chen, Jie Zhao, and Yanyan Li contributed equally to this work. Zifan Chen, Jie Zhao, Yanyan Li, Yang Chen, and Li Zhang directly accessed and verified the underlying data reported in the manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgements\u003c/h2\u003e \u003cp\u003eThe authors sincerely thank all the participants involved in this study for their contribution.\u003c/p\u003e\u003ch2\u003eData availability\u003c/h2\u003e \u003cp\u003eThe datasets will be available upon reasonable request emailed to the corresponding author ([email protected]). The code for the model development is available in Supplementary Materials.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eJoshi, S. S. \u0026amp; Badgwell, B. D. Current treatment and recent progress in gastric cancer. \u003cem\u003eCA Cancer J Clin\u003c/em\u003e \u003cstrong\u003e71\u003c/strong\u003e, 264-279 (2021).\u003c/li\u003e\n\u003cli\u003eMiller, K. D.\u003cem\u003e et al.\u003c/em\u003e Cancer treatment and survivorship statistics, 2022. \u003cem\u003eCA Cancer J Clin\u003c/em\u003e \u003cstrong\u003e72\u003c/strong\u003e, 409-436 (2022).\u003c/li\u003e\n\u003cli\u003eBang, Y. J.\u003cem\u003e et al.\u003c/em\u003e Adjuvant capecitabine and oxaliplatin for gastric cancer after D2 gastrectomy (CLASSIC): a phase 3 open-label, randomised controlled trial. \u003cem\u003eLancet\u003c/em\u003e \u003cstrong\u003e379\u003c/strong\u003e, 315-321 (2012).\u003c/li\u003e\n\u003cli\u003eAl-Batran, S. E.\u003cem\u003e et al.\u003c/em\u003e Perioperative chemotherapy with fluorouracil plus leucovorin, oxaliplatin, and docetaxel versus fluorouracil or capecitabine plus cisplatin and epirubicin for locally advanced, resectable gastric or gastro-oesophageal junction adenocarcinoma (FLOT4): a randomised, phase 2/3 trial. \u003cem\u003eLancet\u003c/em\u003e \u003cstrong\u003e393\u003c/strong\u003e, 1948-1957 (2019).\u003c/li\u003e\n\u003cli\u003eBang, Y. J.\u003cem\u003e et al.\u003c/em\u003e Trastuzumab in combination with chemotherapy versus chemotherapy alone for treatment of HER2-positive advanced gastric or gastro-oesophageal junction cancer (ToGA): a phase 3, open-label, randomised controlled trial. \u003cem\u003eLancet\u003c/em\u003e \u003cstrong\u003e376\u003c/strong\u003e, 687-697 (2010).\u003c/li\u003e\n\u003cli\u003eJanjigian, Y. Y.\u003cem\u003e et al.\u003c/em\u003e Pembrolizumab plus trastuzumab and chemotherapy for HER2-positive gastric or gastro-oesophageal junction adenocarcinoma: interim analyses from the phase 3 KEYNOTE-811 randomised placebo-controlled trial. \u003cem\u003eLancet\u003c/em\u003e \u003cstrong\u003e402\u003c/strong\u003e, 2197-2208 (2023).\u003c/li\u003e\n\u003cli\u003eYu, Y. Y.\u003cem\u003e et al.\u003c/em\u003e The pivotal application of patient-derived organoid biobanks for personalized treatment of gastrointestinal cancers. \u003cem\u003eBiomark Res\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, 73 (2022).\u003c/li\u003e\n\u003cli\u003eDagogo-Jack, I. \u0026amp; Shaw, A. T. Tumour heterogeneity and resistance to cancer therapies. \u003cem\u003eNat Rev Clin Oncol\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 81-94 (2018).\u003c/li\u003e\n\u003cli\u003eDi Capua, D., Bracken-Clarke, D., Ronan, K., Baird, A. M. \u0026amp; Finn, S. The Liquid Biopsy for Lung Cancer: State of the Art, Limitations and Future Developments. \u003cem\u003eCancers (Basel)\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e (2021).\u003c/li\u003e\n\u003cli\u003eDeng, Z., Wu, S., Wang, Y. \u0026amp; Shi, D. Circulating tumor cell isolation for cancer diagnosis and prognosis. \u003cem\u003eEBioMedicine\u003c/em\u003e \u003cstrong\u003e83\u003c/strong\u003e, 104237 (2022).\u003c/li\u003e\n\u003cli\u003eCoombes, R. C.\u003cem\u003e et al.\u003c/em\u003e Personalized Detection of Circulating Tumor DNA Antedates Breast Cancer Metastatic Recurrence. \u003cem\u003eClin Cancer Res\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e, 4255-4263 (2019).\u003c/li\u003e\n\u003cli\u003eStergiopoulou, D.\u003cem\u003e et al.\u003c/em\u003e Comprehensive liquid biopsy analysis as a tool for the early detection of minimal residual disease in breast cancer. \u003cem\u003eSci Rep\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 1258 (2023).\u003c/li\u003e\n\u003cli\u003ePei, H., Li, L., Han, Z., Wang, Y. \u0026amp; Tang, B. Recent advances in microfluidic technologies for circulating tumor cells: enrichment, single-cell analysis, and liquid biopsy for clinical applications. \u003cem\u003eLab Chip\u003c/em\u003e \u003cstrong\u003e20\u003c/strong\u003e, 3854-3875 (2020).\u003c/li\u003e\n\u003cli\u003eConnal, S.\u003cem\u003e et al.\u003c/em\u003e Liquid biopsies: the future of cancer early detection. \u003cem\u003eJ Transl Med\u003c/em\u003e \u003cstrong\u003e21\u003c/strong\u003e, 118 (2023).\u003c/li\u003e\n\u003cli\u003eEdsjo, A.\u003cem\u003e et al.\u003c/em\u003e Precision cancer medicine: Concepts, current practice, and future developments. \u003cem\u003eJ Intern Med\u003c/em\u003e \u003cstrong\u003e294\u003c/strong\u003e, 455-481 (2023).\u003c/li\u003e\n\u003cli\u003eHe, S.\u003cem\u003e et al.\u003c/em\u003e New horizons in the identification of circulating tumor cells (CTCs): An emerging paradigm shift in cytosensors. \u003cem\u003eBiosens Bioelectron\u003c/em\u003e \u003cstrong\u003e203\u003c/strong\u003e, 114043 (2022).\u003c/li\u003e\n\u003cli\u003eChiang, P. J.\u003cem\u003e et al.\u003c/em\u003e Programmed Cell Death Ligand 1 Expression in Circulating Tumor Cells as a Predictor of Treatment Response in Patients with Urothelial Carcinoma. \u003cem\u003eBiology (Basel)\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e (2021).\u003c/li\u003e\n\u003cli\u003eTroncarelli Flores, B. C.\u003cem\u003e et al.\u003c/em\u003e Molecular and Kinetic Analyses of Circulating Tumor Cells as Predictive Markers of Treatment Response in Locally Advanced Rectal Cancer Patients. \u003cem\u003eCells\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e (2019).\u003c/li\u003e\n\u003cli\u003eHe, K.\u003cem\u003e et al.\u003c/em\u003e Decoding the glycoproteome: a new frontier for biomarker discovery in cancer. \u003cem\u003eJ Hematol Oncol\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, 12 (2024).\u003c/li\u003e\n\u003cli\u003eLei, Y.\u003cem\u003e et al.\u003c/em\u003e Combined detection of aneuploid circulating tumor-derived endothelial cells and circulating tumor cells may improve diagnosis of early stage non-small-cell lung cancer. \u003cem\u003eClin Transl Med\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, e128 (2020).\u003c/li\u003e\n\u003cli\u003eRonzoni, M.\u003cem\u003e et al.\u003c/em\u003e Circulating endothelial cells and endothelial progenitors as predictive markers of clinical response to bevacizumab-based first-line treatment in advanced colorectal cancer patients. \u003cem\u003eAnn Oncol\u003c/em\u003e \u003cstrong\u003e21\u003c/strong\u003e, 2382-2389 (2010).\u003c/li\u003e\n\u003cli\u003eLiu, Z.\u003cem\u003e et al.\u003c/em\u003e Combination of tumor markers predicts progression and pathological response in patients with locally advanced gastric cancer after neoadjuvant chemotherapy treatment. \u003cem\u003eBMC Gastroenterol\u003c/em\u003e \u003cstrong\u003e21\u003c/strong\u003e, 283 (2021).\u003c/li\u003e\n\u003cli\u003eMachida, M.\u003cem\u003e et al.\u003c/em\u003e Integrated Predictors by Propensity Scoring With Tumor Markers and Plasma Levels of microRNA-21-5p, IL-17, and IL-10 Complement Early Detection of Hepatocellular Carcinoma in Patients With Liver Cirrhosis. \u003cem\u003eTechnol Cancer Res Treat\u003c/em\u003e \u003cstrong\u003e22\u003c/strong\u003e, 15330338231212084 (2023).\u003c/li\u003e\n\u003cli\u003eSalehi, M.\u003cem\u003e et al.\u003c/em\u003e Circulating Tumor Cells as a Promising Tool for Early Detection of Hepatocellular Carcinoma. \u003cem\u003eCells\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e (2023).\u003c/li\u003e\n\u003cli\u003eWang, H. T.\u003cem\u003e et al.\u003c/em\u003e Circulating tumor cells positivity provides an early detection of recurrence of pancreatic cancer. \u003cem\u003eJ Formos Med Assoc\u003c/em\u003e \u003cstrong\u003e122\u003c/strong\u003e, 653-655 (2023).\u003c/li\u003e\n\u003cli\u003eFeng, Z.\u003cem\u003e et al.\u003c/em\u003e Circulating tumor cells in the early detection of human cancers. \u003cem\u003eInt J Biol Sci\u003c/em\u003e \u003cstrong\u003e18\u003c/strong\u003e, 3251-3265 (2022).\u003c/li\u003e\n\u003cli\u003eVanguri, R. S.\u003cem\u003e et al.\u003c/em\u003e Multimodal integration of radiology, pathology and genomics for prediction of response to PD-(L)1 blockade in patients with non-small cell lung cancer. \u003cem\u003eNat Cancer\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 1151-1164 (2022).\u003c/li\u003e\n\u003cli\u003eBoehm, K. M.\u003cem\u003e et al.\u003c/em\u003e Multimodal data integration using machine learning improves risk stratification of high-grade serous ovarian cancer. \u003cem\u003eNat Cancer\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 723-733 (2022).\u003c/li\u003e\n\u003cli\u003eWang, Q., Li, Z., Zhai, W. \u0026amp; Zheng, J. Clinical values of circulating tumor cells count in localized renal cell carcinoma. \u003cem\u003eTransl Cancer Res\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 2351-2360 (2023).\u003c/li\u003e\n\u003cli\u003eSchoenberg, M. B.\u003cem\u003e et al.\u003c/em\u003e Highly differential count of circulating and tumor infiltrating immune cells in patients with non-HCV/non-HBV hepatocellular carcinoma. \u003cem\u003eCancer Immunol Immunother\u003c/em\u003e \u003cstrong\u003e71\u003c/strong\u003e, 1103-1113 (2022).\u003c/li\u003e\n\u003cli\u003eSyrigos, K., Fiste, O., Charpidou, A. \u0026amp; Grapsa, D. Circulating tumor cells count as a predictor of survival in lung cancer. \u003cem\u003eCrit Rev Oncol Hematol\u003c/em\u003e \u003cstrong\u003e125\u003c/strong\u003e, 60-68 (2018).\u003c/li\u003e\n\u003cli\u003eDevlin, J., Chang, M.-W., Lee, K. \u0026amp; Toutanova, K. Bert: Pre-training of deep bidirectional transformers for language understanding. \u003cem\u003earXiv preprint arXiv:1810.04805\u003c/em\u003e (2018).\u003c/li\u003e\n\u003cli\u003eBrown, T.\u003cem\u003e et al.\u003c/em\u003e Language models are few-shot learners. \u003cem\u003eAdvances in neural information processing systems\u003c/em\u003e \u003cstrong\u003e33\u003c/strong\u003e, 1877-1901 (2020).\u003c/li\u003e\n\u003cli\u003eVaswani, A.\u003cem\u003e et al.\u003c/em\u003e Attention is all you need. \u003cem\u003eAdvances in neural information processing systems\u003c/em\u003e \u003cstrong\u003e30\u003c/strong\u003e (2017).\u003c/li\u003e\n\u003cli\u003eHe, K., Zhang, X., Ren, S. \u0026amp; Sun, J. Deep residual learning for image recognition. In: \u003cem\u003eProceedings of the IEEE conference on computer vision and pattern recognition.\u003c/em\u003e 770-778.\u003c/li\u003e\n\u003cli\u003eDosovitskiy, A.\u003cem\u003e et al.\u003c/em\u003e An image is worth 16x16 words: Transformers for image recognition at scale. \u003cem\u003earXiv preprint arXiv:2010.11929\u003c/em\u003e (2020).\u003c/li\u003e\n\u003cli\u003eYuan, J.\u003cem\u003e et al.\u003c/em\u003e Advanced prompting as a catalyst: Empowering large language models in the management of gastrointestinal cancers. \u003cem\u003eThe Innovation\u003c/em\u003e (2023).\u003c/li\u003e\n\u003cli\u003eAcosta, J. N., Falcone, G. J., Rajpurkar, P. \u0026amp; Topol, E. J. Multimodal biomedical AI. \u003cem\u003eNat Med\u003c/em\u003e \u003cstrong\u003e28\u003c/strong\u003e, 1773-1784 (2022).\u003c/li\u003e\n\u003cli\u003eThirunavukarasu, A. J.\u003cem\u003e et al.\u003c/em\u003e Large language models in medicine. \u003cem\u003eNat Med\u003c/em\u003e \u003cstrong\u003e29\u003c/strong\u003e, 1930-1940 (2023).\u003c/li\u003e\n\u003cli\u003eHuang, Z., Bianchi, F., Yuksekgonul, M., Montine, T. J. \u0026amp; Zou, J. A visual-language foundation model for pathology image analysis using medical Twitter. \u003cem\u003eNat Med\u003c/em\u003e \u003cstrong\u003e29\u003c/strong\u003e, 2307-2316 (2023).\u003c/li\u003e\n\u003cli\u003eTheodoris, C. V.\u003cem\u003e et al.\u003c/em\u003e Transfer learning enables predictions in network biology. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e618\u003c/strong\u003e, 616-624 (2023).\u003c/li\u003e\n\u003cli\u003eRajpurkar, P., Chen, E., Banerjee, O. \u0026amp; Topol, E. J. AI in health and medicine. \u003cem\u003eNat Med\u003c/em\u003e \u003cstrong\u003e28\u003c/strong\u003e, 31-38 (2022).\u003c/li\u003e\n\u003cli\u003eJohnson, K. B.\u003cem\u003e et al.\u003c/em\u003e Precision Medicine, AI, and the Future of Personalized Health Care. \u003cem\u003eClin Transl Sci\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 86-93 (2021).\u003c/li\u003e\n\u003cli\u003eHe, J.\u003cem\u003e et al.\u003c/em\u003e The practical implementation of artificial intelligence technologies in medicine. \u003cem\u003eNat Med\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e, 30-36 (2019).\u003c/li\u003e\n\u003cli\u003eSong, Z.\u003cem\u003e et al.\u003c/em\u003e Clinically applicable histopathological diagnosis system for gastric cancer detection using deep learning. \u003cem\u003eNat Commun\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 4294 (2020).\u003c/li\u003e\n\u003cli\u003eHe, M.\u003cem\u003e et al.\u003c/em\u003e Associations of subcutaneous fat area and Systemic Immune-inflammation Index with survival in patients with advanced gastric cancer receiving dual PD-1 and HER2 blockade. \u003cem\u003eJ Immunother Cancer\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e (2023).\u003c/li\u003e\n\u003cli\u003eHe, M.\u003cem\u003e et al.\u003c/em\u003e Deep learning model based on multi-lesion and time series CT images for predicting the benefits from anti-HER2 targeted therapy in stage IV gastric cancer. \u003cem\u003eInsights Imaging\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 59 (2024).\u003c/li\u003e\n\u003cli\u003eLi, J.\u003cem\u003e et al.\u003c/em\u003e CT-based delta radiomics in predicting the prognosis of stage IV gastric cancer to immune checkpoint inhibitors. \u003cem\u003eFront Oncol\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 1059874 (2022).\u003c/li\u003e\n\u003cli\u003eLu, L., Dercle, L., Zhao, B. \u0026amp; Schwartz, L. H. Deep learning for the prediction of early on-treatment response in metastatic colorectal cancer from serial medical imaging. \u003cem\u003eNat Commun\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 6654 (2021).\u003c/li\u003e\n\u003cli\u003eXu, Y.\u003cem\u003e et al.\u003c/em\u003e Deep Learning Predicts Lung Cancer Treatment Response from Serial Medical Imaging. \u003cem\u003eClin Cancer Res\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e, 3266-3275 (2019).\u003c/li\u003e\n\u003cli\u003eLi, Y.\u003cem\u003e et al.\u003c/em\u003e Evolutionary expression of HER2 conferred by chromosome aneuploidy on circulating gastric cancer cells contributes to developing targeted and chemotherapeutic resistance. \u003cem\u003eClinical Cancer Research\u003c/em\u003e \u003cstrong\u003e24\u003c/strong\u003e, 5261-5271 (2018).\u003c/li\u003e\n\u003cli\u003eLi, Y.\u003cem\u003e et al.\u003c/em\u003e Dynamic monitoring of circulating tumour cells to evaluate therapeutic efficacy in advanced gastric cancer. \u003cem\u003eBritish journal of cancer\u003c/em\u003e \u003cstrong\u003e114\u003c/strong\u003e, 138-145 (2016).\u003c/li\u003e\n\u003cli\u003eNg, A. Y. Preventing\u0026quot; overfitting\u0026quot; of cross-validation data. In: \u003cem\u003eICML.\u003c/em\u003e 245-253.\u003c/li\u003e\n\u003cli\u003eSelvaraju, R. R.\u003cem\u003e et al.\u003c/em\u003e Grad-cam: Visual explanations from deep networks via gradient-based localization. In: \u003cem\u003eProceedings of the IEEE international conference on computer vision.\u003c/em\u003e 618-626.\u003c/li\u003e\n\u003cli\u003ePaszke, A.\u003cem\u003e et al.\u003c/em\u003e Pytorch: An imperative style, high-performance deep learning library. \u003cem\u003eAdvances in neural information processing systems\u003c/em\u003e \u003cstrong\u003e32\u003c/strong\u003e (2019).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-5181858/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5181858/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eGastric cancer (GC) presents challenges in predicting treatment responses due to patient-specific heterogeneity. Recently, liquid biopsies have emerged as a valuable data modality, providing essential cellular and molecular insights and facilitating the capture of time-sensitive information. This study aimed to harness artificial intelligence (AI) technology to analyze longitudinal liquid biopsy data. We collected a dataset from longitudinal liquid biopsies of 91 patients at Peking Cancer Hospital, spanning from July 2019 to April 2022, including 1,895 tumor-related cellular images and 1,698 tumor marker indices. Subsequently, we introduced a Dynamic-Aware Model (DAM) to predict GC treatment responses. DAM incorporates dynamic data through AI components for in-depth longitudinal analysis. Using three-fold cross-validation, DAM exhibited superior performance in predicting treatment responses compared to traditional methods (AUCs: 0.807 vs. 0.582), maintained stable efficacy in the test set (AUC: 0.802), and accurately predicted responses from early treatment data. Moreover, DAM's visual analysis of attention mechanisms identified six key visual features associated strongly with treatment responses. These findings represent a pioneering effort in applying AI technology for interpreting longitudinal liquid biopsy data and employ visual analytics in GC, offering a promising avenue toward precise response prediction and tailored treatment strategies for patients with GC.\u003c/p\u003e","manuscriptTitle":"Predicting response to patients with gastric cancer via dynamic-aware model with longitudinal liquid biopsy data","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-12-03 14:26:34","doi":"10.21203/rs.3.rs-5181858/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"9c8d1cd1-de06-4145-a83f-84ca82fb7e60","owner":[],"postedDate":"December 3rd, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-01-08T22:12:25+00:00","versionOfRecord":[],"versionCreatedAt":"2024-12-03 14:26:34","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-5181858","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5181858","identity":"rs-5181858","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}