美国研究人员开发了新一代面部健康识别模型，能够根据患者的面部特征估算其生物学年龄，并直接评估死亡风险。研究结果显示，这些面部图像中包含与癌症患者预后相关的信息，未来或可帮助医生更全面地评估患者的身体状况，为治疗方案的制定提供参考。
这项名为 FAHR-Face 的基础人工智能模型由美国马萨诸塞州综合医疗机构 Mass General Brigham 的研究团队开发，使用了超过4000万张不同年龄和族群的面部照片进行训练。研究人员以此为基础，开发出两种互补的分析工具：FAHR-FaceAge和FAHR-FaceSurvival。前者估算一个人的生物学年龄，并利用这一信息辅助评估癌症预后；后者则直接从面部照片中预测癌症患者的生存风险，而不是先估算年龄再间接推断结果。
研究团队利用癌症患者的临床照片及其后续医疗结果训练模型。其中，FAHR-FaceSurvival使用了34363张癌症患者面部照片进行针对性训练，并在独立患者数据集中接受验证。研究人员希望通过这种方式，让人工智能从面部图像中识别与患者身体状况和生存结局有关的信号。
研究结果显示，FAHR-FaceSurvival具有较强的风险区分能力。按照模型预测结果，处于死亡风险最高10%区间的患者中，约有一半在90天内死亡；而处于风险最低10%区间的患者中，这一比例约为2%。如果按照四分位数比较，最高风险组的死亡风险也显著高于最低风险组，调整后的风险比为3.23。
研究人员还发现，人工智能估算的生物学年龄与实际年龄之间的差距，同样与癌症患者的生存结局存在关联。在接受放射治疗的患者中，面部看起来比实际年龄大至少5岁的人，在平均约4年的随访期间，死亡风险比其他患者高21%；而面部看起来比实际年龄年轻至少5岁的人，死亡风险则低17%。
这并不意味着一个人看起来年轻，就一定能够更好地抵抗癌症。研究中的生物学年龄是一种由人工智能根据外貌估算的指标，并非直接测量细胞衰老、器官功能或免疫系统状态。它可能综合反映身体消耗、肌肉和软组织变化以及其他与健康有关的因素，但模型究竟识别了哪些具体生理信号，目前仍未完全明确。
研究人员通过分析模型的注意力区域，发现两种工具关注的面部位置有所不同。FAHR-FaceAge主要分析鼻翼至嘴角的法令纹、额头以及太阳穴区域；FAHR-FaceSurvival则更多关注眼睛下方和鼻梁周围。这说明，预测生物学年龄与预测生存风险可能依赖不同的面部信息。
研究团队认为，这项技术未来可能帮助医生在决定是否采用手术、放射治疗或全身性药物治疗之前，更好地评估患者的身体储备和虚弱程度。对于年龄较大、身体状况复杂，或者缺乏完整传统生物标志物数据的患者，面部照片分析有望成为一种成本较低、操作简便的辅助评估方式。
研究还发现，生活方式与面部呈现出的生物学年龄存在关联，其中吸烟与面部显得更衰老之间的关联最为突出，过量饮酒和药物使用也与类似现象有关。研究人员同时表示，FAHR-FaceAge在不同拍摄角度、光线、表情和化妆条件下表现出一定稳定性，在测试过的大多数整形手术后也保持了相对稳定的估算结果，但隆鼻手术会带来较小变化。
尽管如此，研究人员和外部专家都强调，这项技术目前还不能直接取代医生的判断。模型预测的准确性并不自动意味着它能改善实际治疗效果。未来研究仍需确认，将面部分析结果加入医生已经掌握的病史、检查数据和体能评估后，能否进一步提高预测准确性，并真正帮助患者获得更好的治疗结果。
此外，模型训练所用的患者群体主要来自高收入国家，如何确保其适用于不同族群、年龄层和医疗环境，仍是重要问题。由于面部特征可能与种族背景及既往医疗条件有关，人工智能也可能继承训练数据中的偏差。因此，在大规模应用之前，还需要开展更广泛的独立验证和前瞻性临床研究。
相关研究成果已发表于《柳叶刀·数字健康》（The Lancet Digital Health）。研究团队希望，未来能够从一张普通照片中提取更多可量化的健康指标，让面部分析成为临床风险评估的补充工具，而不是依据外貌单独决定患者的治疗机会。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New-generation facial health recognition models have been developed by United States researchers to estimate the biological age of patients based on facial characteristics and directly assess the risk of death. The results of the study show that these facial images contain information relevant to the prognosis of cancer patients and may in the future help doctors to assess the patient ' s state of health more comprehensively and to inform the development of treatment programmes.
This basic artificial intelligence model, known as FAHR-Face, was developed by the research team of Mass General Brigham, an integrated medical facility in Massachusetts, using over 40 million facial photographs of different ages and ethnic groups. On this basis, researchers have developed two complementary analytical tools: FAHR-FaceAge and FAHR-FaceSurvival. The former estimates the biological age of a person and uses this information to support the assessment of cancer projections, while the latter predicts the survival risks of cancer patients directly from facial photographs, instead of estimating age before extrapolating indirectly.
Research teams use clinical photographs of cancer patients and their follow-up medical results training models. Among them, FAHR-FaceSurvival used 34363 facial photographs of cancer patients for targeted training and was validated in the independent patient data collection. In this way, researchers want artificial intelligence to recognize signals related to the patient ' s state of health and fate from facial images.
The results of the study show that the FAHR-FaceSurvival has a strong risk distinction. According to model projections, about half of those at highest risk of death in 10 per cent of districts die within 90 days, while about 2 per cent of those at lowest risk in 10 per cent of districts. The highest risk group also has a significantly higher risk of death compared to the lowest risk group, with an adjusted risk ratio of 3.23.
Researchers have also found that the gap between the biological age of artificial intelligence estimates and the actual age is also linked to the fate of cancer patients. Among patients receiving radiotherapy, the face appears to be at least 5 years older than the actual age, and the risk of death is 21 per cent higher than that of other patients during an average of about four years of follow-up; while the face appears to be at least 5 years younger than the actual age, the risk of death is 17 per cent lower.
This doesn't mean that a person who looks young must be better able to resist cancer. The biological age in research is an indicator estimated by artificial intelligence based on appearance and does not directly measure cell ageing, organ function or immune system state. It may reflect a combination of body consumption, muscle and soft tissue changes and other health-related factors, but it is still unclear what specific biological signals the model identifies.
By analysing the focus areas of the model, researchers have found that the two tools focus on different facial locations. The FAHR-FaceAge mainly analyzes the lines, foreheads and areas of the temple; the FAHR-FaceSurvival pays more attention to the lower eye and the surroundings of the nose. This suggests that predicting biological age and predicting survival risks may depend on different facial information.
The team believes that this technology may in the future help doctors to better assess the physical stock and weakness of patients before deciding whether to use surgical, radiotherapy or whole-body medication. Face photo analysis is expected to be a low-cost, operational and ancillary assessment for patients of older age, complex physical condition or lack of complete data on traditional biomarkers.
The study also found a correlation between lifestyle and the age of biology presented in the face, where smoking and facial ageing were most prominent, with overdrinking and drug use associated with similar phenomena. At the same time, the researchers indicated that the FAHR-FaceAge showed some stability at different angles of filming, light, expression and makeup, and that the estimates had remained relatively stable after most of the plastic surgery tested, but that the nasal surgery had led to minor changes.
Nevertheless, both researchers and outside experts have stressed that this technology cannot currently directly substitute for the doctor ' s judgement. The accuracy of model predictions does not automatically mean that it improves the actual treatment. Future research still needs to confirm whether the incorporation of facial analysis into the medical history, examination data and physical assessment already available to doctors can further improve the accuracy of projections and truly help patients to obtain better treatment outcomes.
In addition, the group of patients used for model training, mainly from high-income countries, continues to be an important issue in ensuring their applicability to different ethnic groups, age groups and medical settings. Artificial intelligence may also inherit deviations from training data, as facial characteristics may be related to ethnic background and medical conditions. There is therefore a need for broader independent validation and forward-looking clinical research before large-scale application.
The results were published in The Lancet Digital Health. The team hopes to be able to extract more quantization health indicators from a common picture in the future, making facial analysis a complementary tool for clinical risk assessment, rather than determining the treatment opportunities of patients on their own, based on their appearance.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.