The validation of orthodontic artificial intelligence systems that perform orthodontic diagnoses and treatment planning
The validation of orthodontic artificial intelligence systems that perform orthodontic diagnoses and treatment planning
复制标题
执行正畸诊断和治疗计划的正畸人工智能系统的验证
DOI：
10.1093/ejo/cjab083
复制
发表时间：
2022
影响因子：
2.6
通讯作者：
Yamashiro Takashi
中科院分区：
文献类型：
--
作者：
Shimizu Yuujin;Tanikawa Chihiro;Kajiwara Tomoyuki;Nagahara Hajime;Yamashiro Takashi
Aim本研究旨在评估两个人工智能（AI）系统，创建一个优先的问题列表和治疗计划，并检查上述系统的性能是否等同于prostitists.Materials和methodsA共967个连续的情况下[800：培训; 67：验证; 100：评价（40：随机选择的临床评价）]。我们使用了一个存储的文档，该文档描述了（1）患者的临床信息，（2）优先级列表，以及（3）没有数字牙齿移动的治疗策略。根据词袋方法（V）对（1）的句子进行向量化;（2）和（3）的句子分别用423和330个标签重新标记。分别使用支持向量机和自注意力网络开发了基于向量V输出优先列表（子任务1）和治疗计划（子任务2）标签的AI系统，同时对系统进行了训练以提高精度和召回率。临床评价由四名儿科医生进行（没有教师或居民;同行小组）在两个会议：在第一个会议上，同行小组和开发的人工智能系统创建问题列表和治疗计划;在第二阶段，两个同行小组（非认可机构）评估了这些清单和计划，包括认可机构的清单和计划，通过使用4分制[不可接受（1）到理想（4）]对其进行评分。结果系统训练后，子任务1和2的准确率分别为65%和48%，召回率分别为55%和48%。子任务1的AI系统的临床评价显示为中等。对于子任务2，AI系统有一个显着低于三个面板的分数，但相同的排名与一个panel.ConclusionsTwo AI系统，输出一个优先的问题列表，并创建一个治疗计划的开发。前者的临床系统能力在同行中处于中等水平，后者的临床系统能力几乎相当于最差的同行。
AimThis study was aimed to evaluate two artificial intelligence (AI) systems that created a prioritized problem list and treatment plan, and examine whether the performance of the aforementioned systems was equivalent to orthodontists.Materials and methodsA total of 967 consecutive cases [800: training; 67: validation; 100: evaluation (40: randomly selected for the clinical evaluation)] were used. We used a stored document that describes (1) the patient’s clinical information, (2) the prioritized list, and (3) a treatment strategy without digital tooth movement. Sentences of (1) were vectorized according to the bag of words method (V); sentences of (2) and (3) were relabelled with 423 and 330 labels, respectively. AI systems that output labels for the prioritized list (subtask 1) and treatment planning (subtask 2) based on the vectorsVwere developed using a support vector machine and self-attention network, respectively, while the system was trained to improve precision and recall. Clinical evaluations were conducted by four orthodontists (no faculty or residents; peer group) in two sessions: in the first session, peer group and the developed AI systems created problem lists and treatment plans; in the second session, two of the peer group (not AI) evaluated these lists and plans, including the lists and plans of the AIs, by scoring them using 4-point scales [unacceptable (1) to ideal (4)]. Scores were compared among the system and peer group (Wilcoxon signed-rank test,P< 0.05).ResultsThe precision after system training was 65% and 48% for subtasks 1 and 2 respectively, with recall of 55% and 48%, respectively. The clinical evaluation of the AI system for subtask 1 showed a mid-rank. For subtask 2, the AI system had a significantly lower score than the three panels but the same rank with one panel.ConclusionsTwo AI systems that output a prioritized problem list and create a treatment plan were developed. The clinical system ability of the former system showed a mid-rank in the peer group, and the latter system was almost equivalent to the worst orthodontist.