阶段	方法与手段	环节	目的
需求分析与方案设计	- 文献研究：查阅知网、学术论文、行业报告，了解痛点与现有方案。 - 用户访谈/问卷：与目标用户（老年人、照护者）沟通，收集真实需求和使用习惯。	研究学习阶段	明确项目核心问题、目标用户、核心功能和市场空白，为解决方案提供理论依据。
软件功能实现	- AI大模型：利用大语言模型进行编程指导、代码生成和调试。 - Android Studio：安卓应用开发环境。 - 计算机视觉技术：使用AI模型实现图像识别和OCR。 - 语音合成（TTS）：将文字信息转换为语音播报。 - 安卓通知服务：实现多维度用药提醒。	软件核心功能	实现智能拍照识别、用药提醒、保质期管理等软件核心功能，确保可行性与用户体验。
硬件功能实现	- Arduino / ESP32 芯片：作为微控制器，控制药格、灯光。 - 舵机/电机：驱动药格自动开启。 - LED灯：提供视觉指引。 - 蓝牙模块：实现手机与药箱的无线通信。 - 各类传感器：如光线、重量传感器等。	硬件核心功能	搭建智能药箱实体，实现药格自动控制、灯光提示以及与手机APP的联动。
软硬件联动与系统集成	- 蓝牙通信协议：建立手机APP与芯片的稳定连接。 - 数据同步机制：实现APP与药箱状态的实时同步。	通信与联动 原型系统搭建	确保软件与硬件无缝协作，形成完整的智能用药生态系统。
原型系统搭建与验证	- AI辅助代码生成：利用DeepSeek 快速生成模块代码。 - 分步式开发：逐步增加功能并测试。 - 用户体验测试：邀请目标用户进行小范围测试，收集反馈。	原型系统搭建	验证设计思路的可行性，发现潜在问题，为后续正式开发提供经验。

在人口老龄化日益加剧的今天，老年群体的健康管理成为一个重要的社会议题。我们通过日常生活的深入观察发现，老年人在日常用药过程中面临着一系列挑战：

• 视觉障碍与用药安全：许多老年人患有老花眼，难以看清药品包装上的小字说明，导致用药信息获取困难，增加了误服或错服的风险。

• 用药依从性挑战：对于需要长期服用多种药物的慢性病患者（如高血压、高血脂、糖尿病等），按时按量服药至关重要。然而，记忆力衰退等因素常导致漏服，严重影响治疗效果，甚至引发健康危机。

• 家庭过期药品管理：许多家庭缺乏有效的过期药品管理机制，过期药物长期存放，不仅占用空间，还可能因误服而对健康造成严重危害。

  我们搜索了知网研究了智能药箱的说明，经过查询发现，本智能药箱的创新点：1、能够自动读取处方服药消息，并自动生成闹铃提示；2、提供放大镜功能，供老人读取小字说明。

​ 为了解决上述问题，我们通过网络搜索，AI大模型问答，头脑风暴，提出一个软硬件结合的智能药箱解决方案。该方案旨在通过整合先进的AI技术与用户友好的界面，为老年人提供一个全方位的用药管理助手。实现如下功能。手机软件功能主要通过Android Studio，利用AI大模型编程实现

​ a. 软件核心功能

• 智能拍照识别：利用手机的摄像头和AI大模型，实现对药品说明书的快速拍照识别。软件能自动提取关键信息，如药品名称、用量、保质期等，并以大字体和语音播报的方式呈现，彻底解决老花眼看不清的问题。
• 多维度用药提醒：通过软件内置的闹钟系统，支持灵活的用药提醒设置。不仅可以根据服药时间定时提醒，还能根据药品名称和用量进行个性化配置，确保用户按时按量服药。
• 保质期智能管理：软件能够自动记录药品的保质期，并在临近过期时发出提醒，有效防止过期药品的误服。

b. 硬件核心功能

• 智能药箱：一个具备智能控制和传感功能的实体药箱。当软件发出用药提醒时，药箱对应的药格会自动打开并亮灯，直观地引导用户取出正确的药物。
• 通信与联动：通过蓝牙连接，实现手机APP与智能药箱的无缝联动，形成一个完整的智能用药生态系统。

​ 通过AI大模型，利用免费的Android Studio，根据我们的设计思路，指挥AI大模型进行相关代码编写出了原型系统。

For building a medical AI agent, you face a critical decision: use a general, off-the-shelf API like a non-medically specific LLM, or fine-tune an open-source model like DeepSeek-7B. There are pros and cons to each approach, and your choice depends heavily on your priorities.

Using a General LLM API Pros:

Ease of Use & Speed: You don't have to manage infrastructure, training, or deployment. This is the fastest way to get a prototype running.

State-of-the-Art Performance: Leading providers like Google, OpenAI, and Anthropic have models with billions or trillions of parameters, offering top-tier reasoning and general knowledge capabilities.

Scalability: These services are designed to handle large numbers of requests, making it easy to scale your app as your user base grows.

Cons:

Lack of Specificity: A general LLM is a broad generalist. It may not understand nuanced medical jargon, specific drug interactions, or hospital protocols. This can lead to less precise and potentially inaccurate responses.

Data Privacy Concerns: Your app will be handling highly sensitive Protected Health Information (PHI). Even if API providers have strong security measures, passing PHI to a third-party service may not comply with regulations like HIPAA or GDPR. You must have explicit legal and technical assurances.

Cost: API calls can be expensive, especially for complex or multi-turn conversations.

Fine-Tuning Your Own LLM (e.g., DeepSeek-7B) Pros:

Domain Specificity: By fine-tuning a model on a curated medical dataset, it becomes a specialist. It will learn to understand medical terminology, reasoning, and specific protocols, leading to more accurate and reliable outputs. This is crucial for a medical application.

Data Control and Privacy: You have complete control over your data pipeline, from training to deployment. This allows you to implement robust security measures and ensure compliance with strict privacy regulations.

Cost-Effectiveness at Scale: While the initial setup for training and hosting is a significant investment, for a high volume of requests, running your own model can become more cost-effective than paying for an API.

Cons:

High Technical Complexity: Building and maintaining a custom AI infrastructure requires specialized knowledge in data engineering, model fine-tuning, and system optimization. You would need a dedicated team.

Resource Intensive: Fine-tuning even a smaller model like DeepSeek-7B requires powerful GPUs and a lot of computing resources.

Performance: A fine-tuned, smaller model may still not perform as well on general reasoning tasks as a large, proprietary model. You would need to carefully evaluate its performance on your specific use cases.

Recommendation For a project like this, where accuracy and data privacy are paramount, fine-tuning an open-source model is the safer and more reliable choice.

While starting with a general API is an option for a quick proof of concept, it's not a viable long-term solution for a medical application due to the critical risks associated with privacy and accuracy. Fine-tuning a model like DeepSeek-7B, possibly using efficient techniques like LoRA (Low-Rank Adaptation), would allow you to create a model that is a true medical expert, not just a generalist. This approach ensures your system can provide medically sound information while maintaining strict control over sensitive patient data.

You can learn about fine-tuning DeepSeek-R1 (a reasoning-focused model) for medical applications in this video. How to Fine-Tune DeepSeek R1 LLM (Step-by-Step Tutorial)