精准精神病学中的算法公平性:临床高风险精神病个体预测模型分析
该研究基于PRONIA研究数据,评估了13个已发表的预测模型(11个预测精神病转化,2个预测功能结局)在临床高风险精神病或近期发作抑郁个体中的公平性,使用准确性平等、预测一致性、假阳性错误率平衡和假阴性错误率平衡等指标,针对性别和受教育程度两个人口学属性进行评估,并与临床医生的判断进行比较。结果显示,预测模型和临床医生判断均存在对低受教育程度个体的系统性偏差,导致11个精神病转化模型中7个的假阳性率更高;算法预测中的偏差模式并不显著比临床医生预测更明显。结论认为,算法和临床医生预测中均存在教育偏差,假设受教育程度较高者结局更有利,这种偏差可能导致低受教育程度患者污名感和社会心理负担增加,而高受教育程度患者的精神病预防不足。
要点
- 预测模型和临床医生判断均对低受教育程度个体分配更不利结局,存在系统性教育偏差。
- 11个精神病转化预测模型中7个在低受教育程度个体中假阳性率更高。
- 算法预测的偏差模式并不显著比临床医生预测更明显。
核心结论
- 算法和临床医生预测中均存在教育偏差,假设受教育程度较高者结局更有利,可能导致低受教育程度患者污名和社会心理负担增加,高受教育程度患者精神病预防不足。
研究设计与样本
- 使用PRONIA研究数据,检查13个已发表模型(11个预测精神病转化,2个预测功能结局)在临床高风险精神病或近期发作抑郁个体中的公平性。
- 使用准确性平等、预测一致性、假阳性错误率平衡和假阴性错误率平衡评估性别和受教育程度相关公平性,并与临床医生判断比较。
主要发现
- 预测模型和临床医生判断均对低受教育程度个体系统性分配更不利结局,导致11个精神病转化模型中7个假阳性率更高。
- 算法预测中的偏差模式并不显著比临床医生预测更明显。
临床怎么用
- 在临床部署预测模型前应评估公平性,避免对特定人口学群体不公平不利,防止延续医疗不平等。
- 需关注低受教育程度患者可能因偏差导致污名和社会心理负担增加,高受教育程度患者可能精神病预防不足。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
BACKGROUND: Computational models offer promising potential for personalised treatment of psychiatric diseases. For their clinical deployment, fairness must be evaluated alongside accuracy. Fairness requires predictive models to not unfairly disadvantage specific demographic groups. Failure to assess model fairness prior to use risks perpetuating healthcare inequalities. Despite its importance, empirical investigation of fairness in predictive models for psychiatry remains scarce. AIMS: To evaluate fairness in prediction models for development of psychosis and functional outcome. METHOD: Using data from the PRONIA study, we examined fairness in 13 published models for prediction of transition to psychosis (n = 11) and functional outcome (n = 2) in people at clinical high risk for psychosis or with recent-onset depression. Using accuracy equality, predictive parity, false-positive error rate balance and false-negative error rate balance, we evaluated relevant fairness aspects for the demographic attributes 'gender' and 'educational attainment' and compared them with the fairness of clinicians' judgements. RESULTS: Our findings indicate systematic bias towards assigning less favourable outcomes to individuals with lower educational attainment in both prediction models and clinicians' judgements, resulting in higher false-positive rates in 7 of 11 models for transition to psychosis. Interestingly, the bias patterns observed in algorithmic predictions were not significantly more pronounced than those in clinicians' predictions. CONCLUSIONS: Educational bias was present in algorithmic and clinicians' predictions, assuming more favourable outcomes for individuals with higher educational level (years of education). This bias might lead to increased stigma and psychosocial burden in patients with lower educational attainment and suboptimal psychosis prevention in those with higher educational attainment.
引用
- Şahin D, Kambeitz-Ilankovic L, Wood S, et al. Algorithmic fairness in precision psychiatry: analysis of prediction models in individuals at clinical high risk for psychosis. Br J Psychiatry. 2024. PMID: 37936347. doi:10.1192/bjp.2023.141
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。