Artificial intelligence (AI) is a transformative technology that has captivated the medical world with its potential to optimize cancer treatment and enhance precision oncology. In cancer diagnosis and treatment, various AI technologies have already provided high-level data examination and analytics that preceding innovations were not capable of. Cancer immunotherapy is a treatment that seeks to boost the immune system to recognize and eradicate tumors. It is a field that is constantly evolving, serving as a fertile environment where AI technologies can accelerate discovery and personalize its regimens. In recent years, AI has played an increased role in the optimization of immunotherapy delivery and drug development. Traditional machine learning and its subfield of deep learning algorithms have already impacted response prediction and related tasks, such as patient stratification for immune checkpoint blockade treatment and identifying potent T-cells in the laboratory to develop effective cellular therapies. Additionally, recently developed technologies such as generative AI (gen AI) and foundation models have expanded upon traditional AI algorithms with new applications such as treatment plan generation and adverse event prediction. As innovations such as agentic AI and the model context protocol (MCP) become increasingly available, efficiency and success in immunotherapy development and delivery could further improve. That said, some challenges must be overcome for AI to reach its full potential in immunotherapy. These include concerns related to data quality control, patient safety, and addressing ethical dilemmas. In this article, we briefly review available state-of-the-art AI technologies for immunotherapy and highlight their capabilities. Then, we examine the current AI applications in immunotherapy including cell therapies, checkpoint inhibitors, and cancer vaccines, covering a diverse array of technologies over a wide range of applications. We analyze the datasets used, performance metrics, and downstream tasks, and highlight existing limitations. Subsequently, we discuss some of the obstacles that have prevented AI from routine clinical adoption. Finally, we envision the future of AI in immunotherapy that may include a framework involving an orchestration of multiple specialized AI agents with a human in the loop.

Advances in medical imaging modalities such as multiparametric and functional MRI, PET, and CT/CBCT, together with complementary innovations such as radiomics, artificial intelligence (AI), adaptive radiotherapy, and theranostics, have expanded the role of imaging from anatomical guidance to biologically informed treatment planning, adaptation, and response assessment. In this review, we provide a comprehensive overview of the technical foundations of imaging biomarkers in radiotherapy (RT), spanning functional and molecular imaging techniques and data-driven analytic approaches. We synthesize current clinical evidence across major disease sites, highlighting how imaging biomarkers are being used to refine target delineation, guide dose painting and functional avoidance, predict and monitor treatment response, and support adaptive and personalized RT strategies. We also critically examine key challenges to clinical translation and implementation, including standardization and reproducibility, validation and generalizability, interpretability of AI-driven models, regulatory and ethical considerations, issues of data sharing, reimbursement, and equity. Finally, we propose a multi-stage translational roadmap to guide the development, validation, and clinical deployment of imaging biomarkers in radiotherapy. Collectively, this review underscores the central role of imaging biomarkers in advancing biologically adaptive and precision radiotherapy and outlines priorities for their responsible and equitable integration into routine clinical practice.

BACKGROUND:Ongoing active monitoring (AM) trials for women with ductal carcinoma in situ (DCIS) are investigating the safety and efficacy of monitoring DCIS lesions vs the current standard of care (surgical treatment). The frequency of upgrade in women undergoing AM for DCIS remains unknown. OBJECTIVE:To evaluate the frequency of upgrade of DCIS at core-needle biopsy to invasive carcinoma at surgical excision among women who meet eligibility criteria for AM trials. METHODS:A retrospective review between 2010 and 2023 was performed of women at an National Cancer Institute-designated comprehensive cancer center with a diagnosis of DCIS at core-needle biopsy who underwent subsequent surgical excision. Medical records were reviewed for clinical presentation, imaging findings, core biopsy, and final surgical pathology. Each patient was evaluated for AM trial eligibility based on published criteria for the LORD, LORIS, and COMET trials. Fisher's exact test compared proportions, with a P-value <.05 considered statistically significant. RESULTS:Of 264 women, 10/264 (3.8%) were eligible for the LORD trial, 24/264 (9.1%) for the LORIS trial, and 64/264 (24.2%) for the COMET trial. Invasive carcinoma was found at surgical excision in 1/10 (10%) patients eligible for the LORD trial, 2/24 (8.3%) for the LORIS trial, and 9/64 (14.1%) for the COMET trial. All occult invasive carcinomas detected at surgical excision in trial-eligible patients were node-negative, with a median size of invasive cancer measuring 3.5 mm (interquartile range, 1-7 mm). CONCLUSION:A subset of women who meet eligibility criteria for DCIS AM trials are at risk for occult invasive carcinoma, with frequency of upgrade ranging from 8% to 14%. CLINICAL IMPACT:More precise criteria and predictive biomarkers are needed to better stratify DCIS lesions and exclude women harboring invasive carcinomas from AM regimens.

OBJECTIVE:To determine whether pre-treatment pelvic bone T1-weighted magnetic resonance imaging provides additional prognostic information beyond traditional clinical variables in patients with cervical cancer undergoing definitive concurrent chemoradiotherapy. METHODS:We retrospectively analyzed 494 treatment-naive patients with cervical cancer who underwent pre-treatment magnetic resonance imaging prior to chemoradiotherapy. Pelvic bone-masked images were processed using a vision transformer framework. A stability-driven feature selection process across repeated Monte Carlo splits and pre-processing configurations identified 2 highly reproducible imaging biomarkers. Survival models were developed using overall survival as the primary endpoint. External validation was performed in 38 patients from The Cancer Genome Atlas without refitting imaging features. The feature-extraction architecture with the highest mean external concordance index was selected, and a representative model was used for clinically interpretable survival analyses. Model-derived 60-month predicted mortality risk was analyzed as a continuous variable in multi-variable Cox proportional hazards models and used for Kaplan-Meier risk stratification. RESULTS:In the discovery cohort, the image-only model achieved a mean concordance index of 0.84 (95% confidence interval 0.82 to 0.85) for overall survival, exceeding that of the clinical-only model (0.52). Imaging-derived risk significantly stratified overall survival in Kaplan-Meier analyses. Five-year overall survival was lower in the high imaging risk group compared with the low imaging risk group within both IB to IIIC1 (69.8% vs 95.4%) and IIIC2 to IVB (47.2% vs 93.7%) stage categories (p <.001 for both). In multi-variable Cox analyses adjusted for age, stage, and histology, imaging-derived risk remained independently associated with overall survival (hazard ratio 1.72 per 10% increase, 95% confidence interval 1.60 to 1.85, p <.001). External validation demonstrated consistent overall survival discrimination (mean concordance index = 0.74). CONCLUSIONS:Pre-treatment pelvic bone magnetic resonance imaging captures reproducible host-related imaging signatures derived from marrow-containing pelvic bone regions that are associated with overall survival and refine conventional stage-based risk stratification in patients with cervical cancer undergoing chemoradiotherapy.

The increasing complexity of medical diagnosis and billing necessitates improving International Classification of Diseases (ICD) coding practices. It also requires partnerships between academia and industry to support rigor and implementation in real-world settings. We conducted an academic-industry collaboration involving four institutions to explore pretrained language models for artificial intelligence (AI)-driven ICD-10 coding in cancer-specific populations. We created a cancer-focused dataset comprising 101,224 clinical notes and 36,040 ICD code assignments across five note types, reflecting real-world scenarios. By fine-tuning and evaluating two existing pretrained language-model approaches designed to process long contexts, PLM-ICD and KEPTLongformer, we developed an institutional benchmark for oncology ICD-10 coding using 20 selected three-character ICD-10 parent codes. The best-performing model achieved an F1-macro of 0.768 and an F1-micro of 0.792 on the test set. Our feasibility analysis showed that integrating these models into clinical workflows could potentially reduce coding time by approximately 10 minutes per case, with a GPU runtime of 7 seconds. Additionally, an exploratory clinical-coder evaluation of attention-based model interpretability showed that 13 of 20 predictions contained high-attention tokens aligned with the corresponding ICD code descriptions; coders rated these tokens as "very helpful" for code assignment. The results demonstrate the potential of AI-driven coding support systems within clinical workflows.

Proton beam therapy (PBT) offers a unique potential for dose conformity to tumors while sparing surrounding healthy tissues. Current PBT accuracy, however, is fundamentally limited by range uncertainties from tissue density variations and anatomical changes, yet no clinically viable methods exist for localizing the dose delivery pulse-by-pulse inside patients during pencil beam scanning (PBS). We developed and clinically demonstrated a first-of-its-kind radiation acoustic beam localization (iRABL) system for real-time tracking PBS trajectory and mapping dose deposition deep in patient's body during PBT. A clinical-grade compact iRABL system featuring high speed, super-resolution, and high sensitivity was specifically designed for PBT applications. Its clinical feasibility was validated through the first-in-human study on prostate cancer patients, demonstrating the capability for in vivo proton dose mapping without interfering with treatment delivery. System performance, including spatial resolution, imaging speed for tracking beam trajectory and temporal dose accumulation, and dosimetric accuracy, was quantitatively characterized using tissue-equivalent phantoms and clinical treatment plans. This iRABL system achieved displacement resolution of 0.1 mm laterally and 0.2 mm axially, exceeding the acoustic diffraction limit by an order of magnitude and surpassing typical proton beam spot sizes. This super-resolution capability, combined with GPU-accelerated image reconstruction and processing, enabled single-pulse detection at a frame rate of 1 kHz, matching the proton system's pulse repetition rate. Dosimetric validation using clinical M-shaped treatment plans met clinical criteria with gamma index passing rates exceeding 90% at 3 mm/3% tolerance, confirming high accuracy for mapping delivered dose distributions. For the first time, by leveraging the high sensitivity and the high speed of our newly developed iRABL system, we are able to localize proton beam and map the proton dose deposition during PBS with sub-diffraction-limit spatial resolution, pulse-by-pulse imaging speed, and clinical grade accuracy. This capability, which addresses fundamental limitations in current treatment monitoring, holds promise for advancing PBT toward image-guided "proton surgery".

Despite recent advances in medical informatics, extracting tumor information from pathology reports remains a challenge in modern cancer registry and surveillance workflows. These documents often have an unstructured format, complex medical content, and a considerably lengthy context, creating significant challenges for automated phenotypic information extraction. Although some recent language models such as BERT, GatorTron, and GPT-4 have demonstrated efficacy in medical applications, they are either constrained by sequence length limitations or cloud-based computing that violates the handling of protected health information. We introduce two oncology pathology-optimized transformer models OncoPT, based on Longformer and BigBird architectures and trained on real-world pathology reports. OncoPT efficiently processes reports up to 4,096 tokens, making it suitable for hospitals’ onsite deployment with limited resources. We apply OncoPT to a common malignancy (exemplified by breast cancer) and a rare malignancy (exemplified by gastric cancer), across five key tumor phenotypes: Subsite, Histology, Grade, Stage, and Laterality. The results demonstrate that OncoPT achieves state-of-the-art weighted F-1 on a private pathology dataset and surpasses commercial chatbots (ChatGPT 4o and o1) on the public CORAL dataset (up to 30% improvement). These findings highlight the robustness of OncoPT models with the added benefit of preserving the privacy of patient health information.

6571 Background: Morphologic diagnosis of myelodysplastic syndromes (MDS) from bone marrow aspirates and biopsies is subjective and error prone. Compared to narrow/specific predictive AI methods used in the literature, recently developed pathology foundation models pretrained on large, diverse histopathology corpora have the potential to improve generalizability while reducing task-specific annotation requirements. Here, we evaluate and compare multiple pretrained foundation models across bone marrow aspirate and biopsy whole-slide images (WSI), incorporate a cell-level bag-of-cells (BoC) foundation model, and assess tri-modal ensemble strategies. Methods: Digitized bone marrow aspirate (Asp) and biopsy (Bx) WSIs from patients with MDS and non-MDS cytopenias were analyzed using four pretrained pathology foundation models (GigaPath, H-Optimus-0, Virchow2-CLS, MUSK) for tile-level feature extraction. Cell-level features from aspirate-derived BoCs were extracted using DinoBloom-B, a foundation model pretrained for hematologic cell representations. Tile- or cell-level features were aggregated using attention-based multiple instance learning (ABMIL). Models were evaluated on 645 patients (319 MDS, 326 non-MDS) from the NHLBI MDS Natural History Study (NCT02775383) using 5-fold cross-validation (CV) and a held-out test set (n=129). For each modality, the best-performing model was selected based on test performance. Predictions were evaluated individually and combined using soft-probability ensembles. An oracle ensemble estimated the upper bound of multimodal complementarity. Results: Performance of the foundation models is summarized in Table 1. A soft-vote tri-modal ensemble of the selected models improved performance beyond any single modality (AUC 0.83; accuracy 0.73). Error analysis demonstrated partially non-overlapping failure modes across modalities, and an oracle ensemble achieved substantially higher accuracy (0.94), indicating unrealized multimodal complementarity. Conclusions: Foundation models enable robust AI-based diagnosis of MDS across bone marrow specimen types. Tri-modal ensembling improves diagnostic performance, while oracle analysis motivates TriPath, a unified framework for joint modeling of aspirate, biopsy, and cellular morphology to support standardized and generalizable MDS diagnosis. Performance summary. Model Bx WSI​ CV AUC Bx WSI​ Test AUC Asp WSI CV AUC Asp WSI Test AUC Asp-derived BoC CV AUC​ Asp-derived BoC Test AUC​ GigaPath+ABMIL​ 0.80+/-0.05 ​ 0.82 ​ 0.77+/-0.06​ 0.66​ ​ H-Optimus-0+ABMIL​ 0.82+/-0.04​ 0.79​ 0.77+/-0.07​ 0.68​ Virchow2-CLS+ABMIL​ 0.81+/-0.04​ 0.76​ 0.78+/-0.05 ​ 0.74 ​ MUSK+ABMIL​ 0.77+/-0.05​ 0.81​ 0.76+/-0.04​ 0.72​ DinoBloom-B+ABMIL​ ​ 0.83+/-0.05 ​ 0.78 ​ Tri-modal Soft Ensemble Test AUC 0.83 Oracle ensemble (upper bound) Test Accuracy 0.94

7521 Background: Real-world ascertainment of clinically meaningful relapse after BCMA CAR-T is challenging because progression signals are distributed across laboratories, imaging, pathology, and clinician actions and are frequently embedded in unstructured external reports. The revised IMWG criteria (Kumar et al., IMS 2025) provide standardized definitions for imaging-based progression (PET/CT and WB-MRI) enabling automated recognition of radiologic and serologic progression thresholds. We developed an AI-enabled multimodal framework to automate real-time derivation of recognized IMWG progression across fragmented care settings. Methods: We analyzed 183 BCMA CAR-T treatment episodes with longitudinal routine laboratory data and independent dual-reviewer adjudication of progression dates. Automated detection evaluated every M-protein and FLC measurement longitudinally in a continuous IMWG rules engine, in addition to flagging new hypercalcemia. In addition, initiation of a new line of therapy (considered a progression event), together with large-language-model extraction of radiologic progression from PET/CT, WB-MRI, CT, X-ray, and MRI brain/spine reports. Performance was assessed using accuracy and specificity. Results: Serologic progression from real-world lab feeds was detected with high reproducibility, with FLC and M-protein achieving high accuracy (96.5% and 98.5%, respectively) and specificity (>97%), indicating minimal premature triggering across serial measurements. Hypercalcemia was rare but highly specific for progression. Radiology-based AI extraction achieved high accuracy (91%) and specificity (92.1%), enabling reliable identification of imaging-defined relapse. Initiation of a new line of therapy occasionally occurred before formal serologic IMWG thresholds were met, reflecting clinician-recognized relapse or that driven by non-serologic disease. Notably, 19% of adjudicated relapses were triggered by radiologic or marrow criteria when serologic IMWG thresholds were absent or lagging, including 12.5% radiology-only and 5.8% marrow-only events, aligned with our published data for post CART relapses (Abuhelwa et al Front. Oncol 2025). Conclusions: Automated lab-only approaches systematically underestimate true IMWG progression after BCMA CAR-T. An AI-enabled multimodal adjudication framework aligned with the revised IMWG criteria enables scalable, real-time, and reproducible progression capture for clinical trials and real-world datasets, supporting rapid endpoint determination, regulatory-grade retrospective analyses, and biologically faithful reconstruction of relapse patterns after CAR-T.

Recent years have witnessed a surge in FDA approved AI tools for healthcare applications. While this growth offers considerable potential benefits for clinical practice, it also introduces substantial challenges related to ethics, regulation, and patient safety. These challenges are further compounded by previously documented gaps in the regulatory approval pathway. These gaps include inconsistent pre-market evaluation practices, over-reliance on retrospective studies, and the limited systematic post-market surveillance of AI devices in real-world clinical settings. Using publicly available FDA data, we developed therefore an interactive web-based dashboard for assessing and predicting the performance of FDA-approved AI software, called PROACTIVE-AI, for the purpose of pro-viding the user with a structured guidance on the anticipated performance of AI-enabled medical devices in real-world clinical settings. The dashboard supports exploratory analysis by diverse stakeholders via knowledge graph visualization and longitudinal trend monitoring of performance indicators, including device recalls and safety-related issues. In addition, PROACTIVE-AI incorporates an AI-aided post-market surveillance risk assessment calculator, derived from historical recall data, to identify device characteristics and con-textual factors associated with elevated deployment risk. Our findings using the PROACTIVE-AI dashboard highlight some of the important challenges related to real-world monitoring and accountability of deployed AI medical devices. Furthermore, it illustrates the potential value of such dashboard in narrowing the trust gap surrounding AI in healthcare by providing quantitative metrics of expected clinical performance and recall-related risk factors.

Purpose: We sent surveys to a large number of radiation oncologists with active thoracic cancer practices and applied the Delphi method over 3 rounds to generate consensus dose-volume histogram metrics. We used these results to create consensus-based organsat-risk dose constraints and target goal templates for practical implementation. Methods and Materials: In this institutional review board-approved study, data were collected using REDCap electronic data capture on a secure server. Radiation oncologists identified from the Accreditation Council for Graduate Medical Education-accredited departments' websites were asked to confirm their self-identification as thoracic radiation oncologists and nominate other respondents. All invitees were asked to complete 3 rounds of questions related to normal tissue constraints, target coverage metrics, prescribing practices, and other planning considerations. Preliminary consensus statements were presented in the second round of surveys for voting on a 5point Likert scale. The third and last round of surveys presented the iterated consensus statements and target coverage metric statements Results: Eighty-three (42.8%) of 194 invitees completed at least 1 round of surveys. The group included a diversity of gender, geography, and clinical settings. Response rates were 83%, 57%, and 55%, respectively, for the 3 rounds. By the end of the process, 48 of 96 (50%) originally proposed normal tissue dose constraint statements were iterated to consensus, and 5 of 7 (71%) proposed target coverage metric statements achieved consensus. These were used to create crowdsourced treatment planning templates. Conclusion: This study achieved broad-based consensus-building on ideal and acceptable dose constraints for conventional, twicedaily, and stereotactic thoracic radiation therapy. Future directions could include extending this approach to other disease sites, studying the influence of widespread implementation on treatment planning, or facilitating the development of community consensus around emergent or controversial questions.

PURPOSE:This study aimed to evaluate the feasibility of using ionizing radiation acoustic imaging (iRAI) to map the delivered dose in patients receiving radiation therapy (RT) with various treatment techniques, including 3-dimensional conformal RT, intensity modulated RT, and volumetric modulated arc therapy. METHODS AND MATERIALS:Patients with intra-abdominal cancer were enrolled in a prospective clinical trial after providing informed consent. Patients were treated with stereotactic body RT using standard clinical techniques with the addition of volumetric iRAI for real-time mapping of 3-dimensional radiation dose deposition. RESULTS:The minimal detectable dose was approximately 10 cGy. The overall shape of the dose distribution and the location of dose deposition, as determined by iRAI, matched the corresponding treatment plan. A gamma passing rate of 75.97% ± 11.12% and 90.99% ± 6.61% with a 10 mm/10% distance to agreement/dose difference suggests good agreement between the iRAI measurement and the treatment plan in both the liver volume and the planning target volume under the current system resolution. A structural similarity value of 0.6284 ± 0.1678 demonstrates that the iRAI measurements have shown structural pattern agreement with the treatment plan within the planning target volume. CONCLUSIONS:This study demonstrates the clinical feasibility of iRAI to monitor radiation dose delivery to deep targets in real time during treatment. This information can be integrated into treatment delivery systems to improve safety and facilitate the adaptation of RT. Despite the promising results achieved with the current form of technology, limitations in detection sensitivity, reconstruction accuracy, and acoustic coupling still need to be addressed.

Large language models (LLMs) show promise in healthcare, but hallucinations remain a major barrier to clinical use. We present CHECK, a continuous-learning framework that integrates structured clinical databases with a classifier grounded in information theory to detect both factual and reasoning-based hallucinations. Evaluated on 1500 questions from 100 pivotal clinical trials, CHECK reduced LLama3.3-70B-Instruct hallucination rates from 31 making an open source model state of the art. Its classifier generalized across medical benchmarks, achieving AUCs of 0.95-0.96, including on the MedQA (USMLE) benchmark and HealthBench realistic multi-turn medical questioning. By leveraging hallucination probabilities to guide GPT-4o's refinement and judiciously escalate compute, CHECK boosted its USMLE passing rate by 5 percentage points, achieving a state-of-the-art 92.1 hallucinations below accepted clinical error thresholds, CHECK offers a scalable foundation for safe LLM deployment in medicine and other high-stakes domains.

PURPOSE:To develop and compare normal tissue complication probability (NTCP) models for recurrent brain metastases (BMs) treated with repeat single-fraction stereotactic radiosurgery (SRS), considering time-dependent discounted prior dose. METHODS AND MATERIALS:We developed three NTCP models (M0, M1-retreat, and M1-combo models) of BMs treated with GammaKnife-based SRS. The maximum dose is 0.2 cc (D0.2cc) of the lesion-specific brain, and the 1-year radionecrosis risk is modeled using a logistic response with doses converted into equivalent dose in 2-Gy fractions (EQD2) based on a linear quadratic linear model. The M0 and M1-retreat models, respectively, predicted radionecrosis risk after SRS to 1029 nonrecurrent lesions (patients, 262) and second SRS to 149 recurrent lesions (patients, 87). The M1-combo model accounted for the second SRS and time-dependent discounted first SRS dose for recurrent lesions estimated using a modified Gompertzian function. RESULTS:All 3 models fitted the data well (χ2, 0.039-0.089, and P = 0.999-1.000). The fitted EQD250 was ∼ 103 Gy for the M0 model, ∼ 88 Gy for the M1-retreat model, and ∼ 165 Gy for the M1-combo model. The fitted γ50 exhibited a progressively flatter dose-response curve across the three models, with values of 1.2 per gray for the M0 model, 0.6 per gray for the M1-retreat model, and 0.4 per gray for the M1-combo model. For the brain D0.2cc of 29 and 19 Gy, the steepest to shallowest dose-response or largest change in NTCP values, ie, NTCP29Gy - NTCP19Gy, was observed in the M1-retreat (0.16), M0 (0.14), and M1-combo (0.06) models. CONCLUSIONS:The model-fitted parameters predicted that recurrent BMs would have a lower threshold dose tolerance and a more gradual dose response to the second SRS than nonrecurrent BMs. This gradual dose-response becomes even more apparent when considering the time-dependent discounted first SRS as a cumulative second SRS. Tailoring SRS retreatment protocols based on NTCP modeling can potentially enhance therapeutic efficacy.

Artificial intelligence (AI) and its machine learning and deep learning algorithms have shown promise in oncological practice. Spatial information analysis in the context of cancer is crucial for its diagnosis and treatment because it can provide an understanding of tumor-microenvironment interactions and reveal insights into response to treatment. AI tools can analyze spatial information at multiple scales, highlighting key disease, clinical, and genetic phenotypes that may reveal underlying mechanisms and molecular markers of response and resistance within the tumor and its microenvironment. By examining tumor interactions at macroscopic (diagnostic imaging) and microscopic (pathology slides and spatial biology) levels, AI can assist in making important diagnostic and prognostic decisions. In this review, we first present an overview of AI and the need for multiscale spatial information in oncology. Then, we examine growing AI applications in the analysis of such information, focusing on diagnostic imaging, digital pathology, and spatial molecular biology. We also discuss applications of large-scale foundation models and task-oriented agentic AI in these fields as emergent technologies. Then, we discuss current limitations for the clinical translation of AI into regular utilization in cancer care and discovery.

Ipilimumab (IPI) improved outcomes for patients with high-risk melanoma compared with IFN-α2b in E1609, a phase III adjuvant trial. We hypothesized that combining candidate immune biomarkers in both tumor and circulating blood could generate a superior predictive biomarker signature. We conducted gene expression profiling on baseline tumors of patients treated with IPI and IFN. We also performed multicolor flow cytometry to compare cellular marker expression on thawed peripheral blood mononuclear cells and Luminex multiplex assay to measure serum biomarkers. We tested the expression levels of 31 genes and 40 circulating biomarkers in relation to survival outcomes. We then developed two separate multivariate Least Absolute Shrinkage and Selection Operator (LASSO) Cox regression models followed by integrative modeling of risk prediction using the prioritized biomarkers. In blood, enriched populations of CXCR3+CD4+ T cells, CXCR3+CD8+ T cells, CTLA4+IFN-γ+CD8+ T cells, and higher levels of CCL3 and CXCL11 were associated with significantly improved overall survival and relapse-free survival, whereas high levels of CTLA4+ regulatory T cells (CD3+CD4+CD25hi+CD152+) and monocytic myeloid-derived suppressor cells (Lin-CD33+HLA-DrloCD14+CD15+) correlated with worse overall survival and relapse-free survival. In tumor, CXCL9, CD8A, CXCL10, and inositol polyphosphate-5-phosphatase D were identified as tier-1 (P < 0.05) and indoleamine 2, 3-dioxygenase 1, Igκ constant, and IL2RB as tier-2 (P < 0.1) biomarkers of survival. Multivariate survival analysis identified that ∼50% of the risk groups were defined by circulating and tumor biomarker models, indicating complementary features of defining risk groups in IPI-treated but not in IFN-treated patients. Integrating candidate blood and tumor immune-related biomarkers generated a baseline signature that maximizes the prediction of immunotherapeutic benefits in reference to the compartmental biomarker signatures.

AI decision support systems can assist clinicians in planning adaptive treatment strategies that can dynamically react to individuals’ cancer progression for effective personalized care. However, AI’s imperfections can lead to suboptimal therapeutics if clinicians over or under rely on AI. To investigate such collaborative decision-making process, we conducted a Human–AI interaction study on response-adaptive radiotherapy for non-small cell lung cancer and hepatocellular carcinoma. We investigated two levels of collaborative behavior: model-agnostic and model-specific; and found that Human–AI interaction is multifactorial and depends on the complex interrelationship between prior knowledge and preferences, patient’s state, disease site, treatment modality, model transparency, and AI’s learned behavior and biases. In summary, some clinicians may disregard AI recommendations due to skepticism; others will critically analyze AI recommendations on a case-by-case basis; clinicians will adjust their decisions if they find AI recommendations beneficial to patients; and clinician will disregard AI recommendations if deemed harmful or suboptimal and seek alternatives. AI decision support system non-homogenously influences clinical decisions in dynamic treatment regimen, as it depends on several factors including prior knowledge, preference, disease type, treatment modality, and AI’s learned behavior and inherent biases.

Brachytherapy is a crucial modality of radiotherapy for cancer, known for its effectiveness in delivering high doses of radiation directly to tumours while sparing surrounding healthy tissues. Despite its clinical importance, recent years have witnessed a concerning decline in its utilization, which negatively impacts patient outcomes. This decline is attributed to several factors, with the inherent complexity of brachytherapy, fair reimbursement policies, and high dexterity being significant barriers. There are silver linings, however, as growing number of applications are seen in continents such as Africa, as well as advances in medical physics technology offering promising solutions to these challenges. This roadmap paper aims to provide a comprehensive overview and preview of advancements in brachytherapy, as well as strategies to address key challenges in four critical areas: 'Imaging and Image Guidance,' 'Treatment Planning,' 'Treatment Delivery,' and 'Brachytherapy Outcomes.' We anticipate that these advances will enhance therapeutic efficacy, equipping clinicians worldwide with the tools needed to deliver state-of-the-art cancer treatments and fostering a promising future in oncology care.

Purpose Patients receiving systemic therapy (ST) for non-small cell lung cancer (NSCLC) experience toxicities that negatively affect patient outcomes. This study aimed to test an approach for prospectively collecting patient-reported outcome (PRO) data, wearable sensor data (WSD), and clinical data, and develop a machine learning (ML) algorithm to predict health care utilization, specifically urgent care (UC) visits. Materials and Methods Patients with NSCLC completed the PROMIS-57 PRO quality-of-life measure and wore a Fitbit to monitor patient-generated health data from ST initiation through day 60. Demographic and clinical data were abstracted from the medical record. ML explainable models on the basis of Bayesian Networks (BNs) were used to develop predictive models for UC visits. Results Patients in the training data set (N = 58) were age 69 years on average (range, 35-89) and mostly female (57%), White (88%), and non-Hispanic (95%) patients with adenocarcinoma (69%). Initial BN models trained on demographic and clinical data demonstrated moderate predictive accuracy on cross-validation for UC visits before ST (AUC, 0.72 [95% CI, 0.57 to 0.80]) and during ST (AUC, 0.81 [95% CI, 0.63 to 0.89]). Incorporating PRO and WSD during ST yielded enhanced models with significantly improved performance (final AUC, 0.86 [95% CI, 0.76 to 0.95]) via DeLong test (P < .001). Conclusion Multidimensional data sources, including demographic, clinical, PRO, and WSD, can enhance ML predictive models to elucidate complex, interactive factors influencing health care utilization during the first 60 days of ST. Use of explainable ML to predict and prevent treatment toxicities and health care utilization could improve patient outcomes and enhance the quality of cancer care delivery.