Identifying and Characterizing Opioid Addiction States Using Social Media Posts
Identifying and Characterizing Opioid Addiction States Using Social Media Posts
复制标题
DOI：
10.1109/bibm52615.2021.9669628
复制
发表时间：
2021-12
期刊：
影响因子：
--
通讯作者：
Deeptanshu Jha;Samantha R. La Marca;Rahul Singh
中科院分区：
文献类型：
--
作者：
Deeptanshu Jha;Samantha R. La Marca;Rahul Singh
阿片成瘾是当代一种重大的健康危机，其复杂性是多种多样的。对任何毒瘾的流行病学建模本身都是具有挑战性的。对于阿片成瘾，由于收集实时数据的困难，以及阿片使用者由于处方滥用相关的耻辱而可能披露的信息的局限性，这一挑战加剧了。在这种情况下，确定个人在(阿片)成瘾阶段的进展是流行病学建模中比较尖锐的问题之一，其解决方案对于在个人和人口层面上设计具体干预措施至关重要。我们描述了一种计算方法，用于从阿片类药物使用者的社交媒体帖子中确定和描述他们的成瘾阶段。该方法结合递归神经网络学习、词联想的信息论分析和基于语境的词嵌入来确定成瘾阶段特定的语言使用情况。使用倾向得分匹配和Logistic回归来识别和描述有很高可能再次吸毒的使用者。实验评估表明，该方法能够区分不同的成瘾阶段，并以较高的准确率识别容易复发的用户，F1得分分别为0.88和0.79。
Opioid addiction constitutes a significant contemporary health crisis that is multifarious in its complexity. Modeling the epidemiology of any addiction is challenging in its own right. For opioid addiction, the challenge is exacerbated due to the difficulties in collecting real-time data and the circumscribed nature of information opioid users may disclose owing to stigma associated with prescription misuse. Given this context, identifying the progression of individuals through the stages of (opioid) addiction is one of the more acute problems in epidemiological modeling whose solution is crucial for designing specific interventions at both personal and population levels. We describe a computational approach for determining and characterizing addiction stages of opioid users from their social media posts. The proposed approach combines recurrent neural network learning with information-theoretic analysis of word-associations and context-based word embedding to determine addiction stage-specific language usage. Users who have a high likelihood for relapsing back to drug-use are identified and characterized using propensity score matching and logistic regression. Experimental evaluations indicate that the proposed approach can distinguish between various addiction stages and identify users prone to relapse with high accuracy as evidenced by F1 scores of 0.88 and 0.79 respectively.