BACKGROUND Separately addressing specific groups of people who share patterns of behavioral change might increase the impact of behavioral interventions to prevent transmission of sexually transmitted infections. We propose a method based on machine learning to assist the identification of such groups among men who have sex with men (MSM). METHODS By means of unsupervised learning, we inferred "behavioral clusters" based on the recognition of similarities and differences in longitudinal patterns of condomless anal intercourse with nonsteady partners (nsCAI) in the HIV Cohort Study over the last 18 years. We then used supervised learning to investigate whether sociodemographic variables could predict cluster membership. RESULTS We identified 4 behavioral clusters. The largest behavioral cluster (cluster 1) contained 53% of the study population and displayed the most stable behavior. Cluster 3 (17% of the study population) displayed consistently increasing nsCAI. Sociodemographic variables were predictive for both of these clusters. The other 2 clusters displayed more drastic changes: nsCAI frequency in cluster 2 (20% of the study population) was initially similar to that in cluster 3 but accelerated in 2010. Cluster 4 (10% of the study population) had significantly lower estimates of nsCAI than all other clusters until 2017, when it increased drastically, reaching 85% by the end of the study period. CONCLUSIONS We identified highly dissimilar behavioral patterns across behavioral clusters, including drastic, atypical changes. The patterns suggest that the overall increase in the frequency of nsCAI is largely attributable to 2 clusters, accounting for a third of the population.

中文翻译：

---

与男人发生性关系的人类免疫缺陷病毒阳性男性的性行为群显示出高度不同的时间趋势。

背景技术单独解决共享行为改变模式的特定人群可能会增加行为干预的效果，以防止性传播感染的传播。我们提出了一种基于机器学习的方法，以帮助在与男性发生性关系的男性中识别此类人群（MSM）。方法通过无监督学习，我们在过去18年的HIV队列研究中，基于对与非稳定伴侣（nsCAI）进行安全套肛门性交的纵向模式的相似性和差异的识别，推断出“行为簇”。然后，我们使用监督学习来调查社会人口统计变量是否可以预测聚类成员。结果我们确定了4个行为簇。最大的行为集群（集群1）占研究人群的53％，并且表现出最稳定的行为。聚类3（占研究人群的17％）显示出持续增加的nsCAI。社会人口统计学变量对于这两个集群都是可预测的。其他两个集群显示出更大的变化：集群2（占研究人口的20％）中的nsCAI频率最初与集群3相似，但在2010年有所加快。集群4（占研究人口的10％）对nsCAI的估计值明显较低直到2017年nsCAI都比其他所有集群都大幅度提高，到研究期结束时达到了85％。结论我们确定了跨行为群的高度不同的行为模式，包括剧烈的，非典型的变化。