大语言模型训练逻辑启发下的预防医学专业任务能力培养

Cultivation of Professional Task Competence in Preventive Medicine Inspired by the Training Logic of Large Language Models

摘要: 生成式人工智能正在改变医学教育中的任务完成方式，但专业能力并不会因知识掌握或案例接触而自动形成。预防医学学生需要在流行病学调查、疾病与危险因素监测、环境与职业健康风险评估、健康教育与人群干预、突发公共卫生事件处置等不同专业任务中反复调用知识，并逐步学会处理信息缺失、资源限制和多重目标。本文借鉴大语言模型在预训练基础上通过任务适配和指令微调形成下游任务能力的设计逻辑，提出预防医学“任务微调式”专业训练框架。该框架以完整专业任务而非知识题目作为训练单位，从任务结构度、信息完整度、决策风险、资源约束和社会复杂性五个维度组织任务分布，并形成“结构化任务–综合任务–开放任务–动态任务”四级训练阶梯。在此基础上，以营养与食品卫生学相关任务作为示例，说明分层任务库如何在具体课程中落地，并讨论生成式人工智能在生成情境变式、提供反例和扩展任务样本方面的辅助作用。本文强调，所谓“任务微调”是教学设计类比，其核心是系统安排学生经历的任务类型和复杂度，使已有知识在不同公共卫生情境中被反复调用、整合和重组，逐步转化为可执行的专业能力。本文将该框架置于胜任力导向教育、可托付专业活动、PBL/CBL与适应性专长等理论脉络中加以定位，阐明其对现有模型在任务复杂度系统编排方面缺口的补充；并进一步提出一套可操作的有效性检验研究设计方案(含研究假设、准实验设计、测量工具与数据分析策略)，为后续实证检验提供路径。

Abstract: Generative artificial intelligence is changing how professional tasks are accomplished in medical education, yet professional competence does not form automatically from mastering knowledge or being exposed to cases. Preventive medicine students need to repeatedly invoke knowledge across diverse professional tasks—epidemiological investigation, disease and risk-factor surveillance, environmental and occupational health risk assessment, health education and population intervention, and emergency public health response—and gradually learn to handle missing information, limited resources, and competing goals. Drawing on the design logic by which large language models, after pre-training, acquire downstream task capabilities through task adaptation and instruction fine-tuning, this paper proposes a “task fine-tuning” professional training framework for preventive medicine. Taking complete professional tasks rather than knowledge items as the unit of training, the framework organizes task distribution along five dimensions—task structuredness, information completeness, decision risk, resource constraints, and social complexity—and forms a four-level training ladder of “structured - integrated - open - dynamic” tasks. Using nutrition- and food-hygiene-related tasks as examples, the paper illustrates how a layered task bank can be implemented in a specific course, and discusses the supporting role of generative AI in producing situational variants, counterexamples, and expanded task samples. The paper emphasizes that “task fine-tuning” is an instructional-design analogy whose core is to systematically arrange the types and complexity of tasks students experience, so that existing knowledge is repeatedly invoked, integrated, and reorganized across public health contexts and gradually transformed into executable professional competence. The paper situates the framework within the theoretical traditions of competency-based education, entrustable professional activities, PBL/CBL, and adaptive expertise, clarifying how it complements the gap these models leave in the systematic sequencing of task complexity; it further proposes an operable research design for testing the framework’s effectiveness (including hypotheses, a quasi-experimental design, measurement instruments, and analytic strategies), thereby providing a pathway for subsequent empirical testing.