Agentic Large Language Models for Healthcare: Current Progress and Future Opportunities

Medicine Advances, Journal Year: 2025, Volume and Issue: unknown

Published: March 3, 2025

Language: Английский

---

Toward real‐world deployment of machine learning for health care: External validation, continual monitoring, and randomized clinical trials

Health care science, Journal Year: 2024, Volume and Issue: 3(5), P. 360 - 364

Published: Oct. 1, 2024

In this commentary, we elucidate three indispensable evaluation steps toward the real-world deployment of machine learning within healthcare sector and demonstrate referable examples for diagnostic, therapeutic, prognostic tasks. We encourage researchers to move beyond retrospective within-sample validation, step into practical implementation at bedside rather than leaving developed models in dust archived literature. Machine (ML) has been increasingly used tackling various tasks owing its capability learn reason without explicit programming [1]. Most ML have had their accuracy proven through internal validation using data. However, external data, continual monitoring prospective randomized controlled trials (RCTs) data are important translation clinical practice [2]. Furthermore, ethics fairness across subpopulations should be considered throughout these evaluations. Different from which evaluates performance a subset original datasets, assesses contexts that may vary subtly or considerably one they were [3]. External serves rectify inflated estimates capabilities overfitting guarantees generalizability transportability diverse populations [4]. For can leverage abundant resources publicly accessible databases such as PhysioNet [5]. Three scenarios recommended after identifying suitable database with sufficient sample size guarantee testing robustness [6]. The first involves directly deploying trained on simulate brand-new scenario previous second entails large training set new fine-tune models, simulating ample collected context [7]. third represents an intermediate situation wherein gradually fed where deployed setting, incrementally collected, updated iteratively newly [8]. existing studies focused direct [9]. Holsbeke et al. [10] previously published diagnostic detecting adnexal mass malignancy multiple medical centers different countries population characteristics. therapeutic pertinent reference is study investigating survival benefits adjuvant therapy breast cancer evaluated originally United Kingdom, settings States [11]. realm tasks, Clift [12] offered comprehensive approach externally validate predicting 10-year risk cancer-related mortality, detailing methods calculation, identification, outcome definition, evaluation. addition assessing model performance, similarity between datasets quantified enable elucidation degradation further identify potential avenues enhancement [13]. Following large-scale subsequent specific setting [14]. Specifically, receive make predictions accordingly, predefined time frame Compared step, distribution drift, control quality, trigger system alarms when deviates normal behavior target [15]. Because operation mainly conducted by professionals, developers focus user-friendly practice. aspect offline hospital allocated computation would limited low latency responding other functions inside system. development secure privacy-aware maintenance method quickly addressing technical collapses while minimizing access patients' private last interface Android app [16] web-based software [17] facilitates use health care professionals comprehends suggestions. It emphasized application designed operate independently from, not interfere with, decision-making processes. This precaution necessary avoid any adverse impact quality. Exemplary seen work Wissel [18]. Those authors prospective, real-time assessment ML-based classifiers epilepsy surgery candidacy Cincinnati Children's Hospital Medical Center. To mitigate risks associated classifiers, patients who deemed appropriate surgical candidates algorithm subjected manual review two expert epileptologists, final decisions confirmed via chart review. A critical insight was effective necessitates synergistic collaboration clinicians, provide essential expertize, information technology contribute research operational knowledge [19, 20]. Assuming tool demonstrates accurate pursue approval RCTs administrative committees. tools classic four-phase RCTs. ensure safety real-life scenarios, absolutely interventions likely avoided. recommend designing compare diagnosis clinicians (intervention group) (control [21-23]. instance, He [24] implemented ML-guided workflows reduced required sonographers cardiologists diagnoses left ventricular ejection fraction. seek ethical institutional board comply standards regulations. Then, proceed Phase I trial assess (whether introduction distracts impairs diagnoses) used. II, few hundred recruited whether statistically significant improvements result clinicians' diagnoses. III, several even thousand effectiveness tool, demonstrating superiority over solutions. If receives agency then investigate wider range IV. Upon efficacy rigorously RCTs, national regulatory agencies US Food Drug Administration (FDA) commercialization [25]. paradigmatic illustration found Titano [26]. three-dimensional convolutional neural networks diagnose acute neurological events head computed tomography images. efficiency subsequently validated randomized, double-blind, trial. suggest referring Nimri [27]. multicenter multinational physicians specialized academic diabetes optimizing insulin pump doses. Mayo Clinic 1-year occurrence asthma exacerbation [28]. detailed guideline conducting could benefit FDA's Policy Device Software Functions Mobile Applications [29], includes provisions applications apply algorithms [30]. Alongside population-level evaluations, there burgeoning awareness about implications revealed diagnose, treat, bill inconsistently [31]. Therefore, it imperative equity patient outcomes, resource allocation [31-33]. Thompson [34] proposed framework biases recalibration modules. module adjusted decision cutoff threshold affected bias, recalibrated outputs, enhancing congruence observed events. Chen [31] systematically summarized path fair medicine, subpopulation collection federated learning, principles, operationalization ecosystems, independent regularization governance disparities. Apart assessments, endorsement thoroughly integrated processes [31, 35]. light these, buried Han Yuan: Conceptualization (lead); curation formal analysis investigation methodology writing—original draft writing—review editing (lead). like acknowledge Prof. Nan Liu Duke-NUS School his invaluable support. author declares no conflict interest. exempt committee because did involve human participants, animal subjects, sensitive collection. Not applicable. Data sharing applicable article generated analyzed during current study.

Language: Английский

---

FLDQN: Cooperative Multi-Agent Federated Reinforcement Learning for Solving Travel Time Minimization Problems in Dynamic Environments Using SUMO Simulation

Sensors, Journal Year: 2025, Volume and Issue: 25(3), P. 911 - 911

Published: Feb. 3, 2025

The increasing volume of traffic has led to severe challenges, including congestion, heightened energy consumption, increased air pollution, and prolonged travel times. Addressing these issues requires innovative approaches for optimizing road network utilization. While Deep Reinforcement Learning (DRL)-based methods have shown remarkable effectiveness in dynamic scenarios like management, their primary focus been on single-agent setups, limiting applicability real-world multi-agent systems. Managing agents fostering collaboration a reinforcement learning scenario remains challenging task. This paper introduces cooperative federated algorithm named FLDQN address the challenge agent cooperation by solving time minimization challenges (MARL) scenarios. leverages facilitate knowledge sharing among intelligent agents, vehicle routing reducing congestion environments. Using SUMO simulator, multiple equipped with deep Q-learning models interact local environments, share model updates via server, collectively enhance policies using unique observations while benefiting from collective experiences other agents. Experimental evaluations demonstrate that achieves significant average reduction over 34.6% compared non-cooperative simultaneously lowering computational overhead through distributed learning. underscores vital impact provides an solution enabling environment.

Language: Английский

---

Designing Deep Reinforcement Learning enhanced edge-terminal collaborative AIoT for Intelligent Visitor Management System

Ad Hoc Networks, Journal Year: 2025, Volume and Issue: unknown, P. 103756 - 103756

Published: Jan. 1, 2025

Language: Английский

---

Materiality and risk in the age of pervasive AI sensors

Nature Machine Intelligence, Journal Year: 2025, Volume and Issue: unknown

Published: March 20, 2025

Language: Английский

---

A vision of human–AI collaboration for enhanced biological collection curation and research

BioScience, Journal Year: 2025, Volume and Issue: unknown

Published: March 28, 2025

Abstract Natural history collections play a crucial role in our understanding of biodiversity, informing research, management, and policy areas such as biosecurity, conservation, climate change, food security. However, the growing volume specimens associated data presents significant challenges for curation management. By leveraging human–AI collaborations, we aim to transform way biological are curated managed, realizing their full potential addressing global challenges. In this article, discuss vision improving management using collaboration. We explore rationale behind approach, faced general problems, benefits that could be derived from incorporating AI-based assistants collection teams. Finally, examine future possibilities collaborations between human digital curators collection-based research.

Language: Английский

---

A framework reforming personalized Internet of Things by federated meta-learning

Nature Communications, Journal Year: 2025, Volume and Issue: 16(1)

Published: April 20, 2025

Language: Английский

---

Scaling Up Multi-Agent Reinforcement Learning: An Extensive Survey on Scalability Issues

IEEE Access, Journal Year: 2024, Volume and Issue: 12, P. 94610 - 94631

Published: Jan. 1, 2024

Language: Английский

---

A Survey on Knowledge Distillation: Recent Advancements

Machine Learning with Applications, Journal Year: 2024, Volume and Issue: 18, P. 100605 - 100605

Published: Nov. 10, 2024

Language: Английский

---

Evidence-based Diagnostic Performance Benchmarks in Prostate MRI: An Unmet Clinical Need

Radiology, Journal Year: 2024, Volume and Issue: 312(2)

Published: Aug. 1, 2024

Language: Английский

---