Personalized AI Model for Antibiotics Prescribing

抗生素广泛用于治疗感染性疾病或预防某些类型的细菌感染。为了确保抗生素治疗的有效性，药敏试验用于支持抗生素的选择。然而，抗生素敏感性测试通常需要2天才能扭转，并且只有当细菌感染可以被确认时（约50%）。由于对个体耐药风险因素的了解也有限，这大大延迟了适当的治疗，通过经验性使用广谱抗生素对抗菌素耐药性产生压力，并增加了患者的不必要费用。此外，需要仔细考虑与使用抗生素相关的许多风险，例如不良副作用，并发症，抗生素耐药性的发展以及与患者正在服用的其他药物的负面相互作用。该项目旨在开发一个个性化的基于机器学习的模型，并根据一个庞大的、相互关联的初级和二级医疗记录数据集（全系统范围的人类发展报告-英国数据集），对不确定性进行阐述，以指导抗生素处方。专注于临床决策支持，我们将训练抗生素耐药性预测模型，并研究哪些可修改的风险因素预测抗生素耐药性以及为什么，耐药性如何沿着时间变化，其他药物是否影响抗生素的有效性以及如何为卫生服务和政策研究提供信息。基于模型和这些问题的答案，我们将启用处方算法，为患者提供个性化的抗生素处方建议。最后，将仔细评估模型和算法，调查医生对此类系统提出的建议的看法，并与临床医生共同设计抗菌管理仪表板。

Antibiotics are widely used for treating patients with infectious diseases or preventing some types of bacterial infections. To ensure the effectiveness of antibiotic therapy, susceptibility tests are used to support antibiotic selection. However, antibiotic susceptibility tests routinely take 2 days to turn around and only when bacterial infection can be confirmed (~50%). As there is also limited understanding of an individual's risk factors for resistance, this substantially delays appropriate therapy, drives pressures on antimicrobial resistance through the empiric use of broad-spectrum antibiotics, and increases unnecessary costs for patients. Additionally, many risks associated with the use of antibiotics need to be carefully considered, such as adverse side effects, complications, the development of antibiotic resistance, and negative interactions with other medications patients are taking. This project aims at developing a personalized machine-learning-based model with the exposition of uncertainty to guide antibiotics prescription, based on a huge, linked dataset of primary and secondary care records, the systemwide HDR-UK dataset. Focusing on clinical decision support, we will train an antibiotic resistance prediction model and investigate what modifiable risk factors predict antibiotic resistance and why, how the resistance changes along with time, whether other medications affect the effectiveness of antibiotics and how, to inform health services and policy research. Based on the model and answers to these questions, we will enable prescribing algorithms that give individualized antibiotic prescription recommendations for patients. Finally, the model and algorithms will be carefully assessed, doctors view about recommendations made by such systems will be investigated, and an antimicrobial stewardship dashboard will be co-designed with clinicians.