An AI-based Multimodal Approach to Predict Pain in Postnatal Care Scenarios

项目摘要 在过去的十年中，技术和外科手术的进步导致了一个显着的 接受救生手术的新生儿人数增加。这些术后新生儿 通常被分流到新生儿重症监护室（NICU）进行疼痛管理， 阿片类药物，主要是吗啡、芬太尼和美沙酮。然而，大量的证据表明，在- 体外，动物和人体研究强烈表明，这种严重的疼痛，阿片类药物方案造成 对发育中的神经系统造成长期和可能永久的创伤性伤害， 新生儿 我们提出了一种新的机器学习和计算机视觉方法用于早期疼痛检测 (EPD)重点是NICU中的术后新生儿。通过提醒NICU护理人员 在疼痛发作前约30分钟，EPD将允许NICU工作人员使用速效，阿片类药物 药物，例如，静脉注射对乙酰氨基酚、布洛芬或酮咯酸，与非 药理学方法，以“保持领先的痛苦”，同时避免阿片类药物治疗， 耐受性、戒断以及相关的副作用。 我们的NICU专家，生物科学家和计算机专家团队将证明- 术后新生儿可靠EPD的概念，目的如下： 1)收集临床信息和多模态数据（面部表情，身体运动， 哭闹频率、生命体征）用于新生儿的术后疼痛预测。 我们将收集和标记多模态信号（面部表情，身体动作，哭泣 频率，生命体征），来自坦帕综合医院NICU的约60例术后疼痛新生儿 医院（TGH）。我们将联合收割机将这些数据与另一个约60名新生儿队列的类似数据相结合 2019年至2022年，在TGH使用相同的系统和方法收集了（总计约120名新生儿）。 2)显示预测新生儿术后疼痛发作的概念验证。 从目标1中的训练案例中收集的多模态数据将为 训练卷积神经网络预测术后新生儿疼痛发作时间 in the testing测试cases案件.测试用例中EPD的性能目标是疼痛预测~ 30 疼痛发作前10分钟，置信概率为90%。我们的目的是扰乱电流 标准[手术;镇静;术后疼痛;阿片类药物依赖、耐受、戒断; 出院]赞成更安全的阿片类药物保留方法[手术;镇静;非阿片类药物治疗; 放电]。我们的2期研究将增加来自更多样化患者人群的数据， EPD对应激生物标志物的可能影响，例如，头发皮肤血液中的皮质醇去甲肾上腺素 或尿液。对公共卫生的主要好处将是保护可能是最脆弱的人 患者群体免受对其未来健康和福祉的不必要损害。

PROJECT SUMMARY Advances in technology and surgical procedures in the past decade have led to a remarkable increase in numbers of newborns subjected to lifesaving surgery. These postoperative neonates are customarily triaged to neonatal intensive care units (NICUs) for pain management with opioids, primarily morphine, fentanyl, and methadone. However, substantial evidence from in- vitro, animal, and human studies strongly suggests this severe pain-to-opioids regimen causes long-lasting and likely permanent traumatic harm to the developing neurological systems of neonates. We propose a novel machine learning and computer vision approach for early pain detection (EPD) with emphasis on postoperative neonates in NICU. By alerting NICU caregivers a minimum of ~ 30 minutes prior to pain onset, EPD will allow NICU staff to use fast-acting, opioid-sparing medications, e.g., intravenous paracetamol, ibuprofen, or ketorolac, in conjunction with non- pharmacological approaches, to “stay ahead of the pain” while avoiding opioid treatments, tolerance, withdrawal, and their associated side effects. Our team of NICU specialists, bioscientists, and computer experts will demonstrate proof-of- concept for reliable EPD in post-surgical neonates using the following aims: 1) Collect clinical information and multimodal data (facial expression, body movement, crying frequency, vital signs) for post-surgical pain prediction in neonates. We will collect and label multi-modal signals (facial expression, body movements, crying frequency, vital signs) from ~ 60 neonates in post-surgical pain at the NICU at Tampa General Hospital (TGH). We will combine these data with similar data from another cohort of ~60 neonates (total ~ 120 neonates) collected using the same system and approach at TGH from 2019 to 2022. 2) Show proof-of-concept for predicting the onset of post-surgical pain in neonates. The multimodal data collected from the training cases in Aim 1 will provide the ground truth for training a convolutional neural network to predict time-to-onset of pain for postoperative neonates in the testing cases. The performance target for the EPD in the test cases is pain prediction ~ 30 minutes prior to pain onset with a 90% confidence probability. Our intention is to disrupt the current standard [surgery; sedation; postoperative pain; opioid dependence, tolerance, withdrawal; discharge] in favor of a safer opioid-sparing approach [surgery; sedation; non-opioid treatment; discharge]. Our Phase 2 studies will add data from more diverse patient populations and examine the possible effects of EPD on stress biomarkers, e.g., cortisol, norepinephrine in hair, skin, blood, or urine. The major benefit to public health will be protection of perhaps the most vulnerable patient populations from unnecessary damage to their future health and well-being.