Table 3.
Summary of generalized large language model–driven intervention design in stroke care.
| Study | Country | Stage in the stroke care continuum | Foundation model or model series | Access | Instruction design | Other adaptation strategies |
| Pedro et al [38] | Portugal | Prognosis | GPT-3.5 | Web-based chat interface (ChatGPT) | Zero shot | None |
| Chen et al [39] | United States | Treatment | GPT-4 | Web-based chat interface (ChatGPT) | Zero shot, role based, context enhanced, and format constrained | None |
| Strotzer et al [40] | Germany | Diagnosis | GPT-4-1106-vision-preview | Official APIa (via OpenAI platform) | Zero shot, role based, context enhanced, and format constrained | None |
| Kuzan et al [41] | Turkey | Diagnosis | GPT-4 Vision | Web-based chat interface (ChatGPT) | Zero shot and context enhanced | None |
| Fei et al [42] | China | Rehabilitation | GPT-3.5 and GPT-4 | Web-based chat interface (ChatGPT) | Zero shot, role based, and context enhanced | None |
| Lee et al [43] | —b | Diagnosis | GPT-4 | Unclarified | Zero shot, chain of thought, context enhanced, and format constrained | None |
| Haim et al [44] | Israel | Diagnosis and treatment | GPT-3.5 and GPT-4 | Web-based chat interface (ChatGPT) | Zero shot and context enhanced | None |
| Chen et al [45] | United States | Diagnosis | GPT-4 | Web-based chat interface (Bing chat) | Zero shot, role based, and context enhanced | None |
| Blacker et al [46] | United States | Treatment | GPT-4 | Web-based chat interface (ChatGPT) | Zero shot and context enhanced | None |
| Zhang et al [37] | Japan | Rehabilitation | GPT-4 | Web-based chat interface (ChatGPT) | Zero shot | None |
| Sivarajkumar et al [47] | United States | Rehabilitation | GPT-3.5-turbo | Official API (via Microsoft Azure) | Zero shot, few shot, role based, and format constrained | None |
| Guo et al [48] | China | Diagnosis and treatment | BARTc-base-Chinese and BART-large-Chinese | Unclarified | — | Fine-tuning, constrained decoding, encoding representation reuse, beam search, feature fusion, and shared encoder weights |
| Lehnen et al [49] | Germany | Treatment | GPT-3.5 and GPT-4 | Web-based chat interface (ChatGPT) | Zero shot, format constrained, and context enhanced | None |
| Fiedler et al [50] | United States | Diagnosis, treatment, prognosis, and rehabilitation | GPT-3.5-turbo-16k | Official API (via Microsoft Azure) | Zero shot, role based, format constrained, and context enhanced | Temperature set to 0 |
| Wang et al [51] | China | Treatment | GPT-3.5-turbo, GPT-4, Gemini Pro, GLM-4, Spark 3, and Qwen-Max | Official APIs (via unclarified platforms) | Zero shot, format constrained, and context enhanced | None |
| Goh et al [52] | Australia | Diagnosis and treatment | Llama 3-70B | Local inference | Zero shot, role based, and format constrained | Temperature set to 0 |
| Baro et al [53] | Brazil | Prevention | openCabrita 3B | Unclarified | — | Low-rank adaptation tuning |
| Meddeb et al [54] | Germany | Treatment | Qwen-72B, Mixtral 8x7B, and BioMistral-7B | Local inference | Zero shot, format constrained, and context enhanced | None |
| Kim et al [55] | South Korea | Diagnosis, treatment, and prognosis | GPT-4-32k | Official API (via unclarified platform) | Few shot, format constrained, and context enhanced | Low-temperature setting |
| Argymbay et al [56] | Canada | Prevention | BioMistral-7B | Private API (via Hugging Face on Amazon SageMaker) | Few shot and context enhanced | Temperature set to 0.3 |
| Neo et al [57] | Singapore | Rehabilitation | GPT-3.5-turbo and PaLM 2 | Web-based chat interfaces (ChatGPT and Google Bard) | Zero shot and context enhanced | None |
| Wu et al [58] | United States | Prevention | GPT-3.5 | Web-based chat interface (ChatGPT) | Zero shot | None |
| Chen et al [59] | China | Rehabilitation | GPT-4, GPT-3.5-turbo, and GLM-130B | Official APIs (via unclarified platforms) | Few shot, role based, format constrained, and context enhanced | None |
| Rifai et al [60] | Indonesia | Rehabilitation | GPT-4o | Official API (via unclarified platforms) | Zero shot, format constrained, and context enhanced | Temperature set to 0.5; token generation minimized |
| Anghelescu et al [36] | — | Treatment | Unclarified GPTd | Web-based chat interface (ChatGPT) | Zero shot | None |
aAPI: application programming interface.
bNot applicable.
cBART: bidirectional and auto-regressive transformers.
dGPT: generative pretrained transformer.