Hutt Suzanna, Mihaies Denis, Karteris Emmanouil, Michael Agnieszka, Payne Annette M, Chatterjee Jayanta
Academic Department of Gynaecological Oncology, Royal Surrey NHS Foundation Trust Hospital, Guildford GU2 7XX, UK.
Department of Clinical and Experimental Medicine, Faculty of Health and Medical Sciences, School of Biosciences and Medicine, University of Surrey, Guildford GU2 7XH, UK.
Cancers (Basel). 2021 Jul 22;13(15):3689. doi: 10.3390/cancers13153689.
In this study we wished to determine the rank order of risk factors for endometrial cancer and calculate a pooled risk and percentage risk for each factor using a statistical meta-analysis approach. The next step was to design a neural network computer model to predict the overall increase or decreased risk of cancer for individual patients. This would help to determine whether this prediction could be used as a tool to decide if a patient should be considered for testing and to predict diagnosis, as well as to suggest prevention measures to patients.
A meta-analysis of existing data was carried out to calculate relative risk, followed by design and implementation of a risk prediction computational model based on a neural network algorithm.
Meta-analysis data were collated from various settings from around the world. Primary data to test the model were collected from a hospital clinic setting.
Data from 40 patients notes currently suspected of having endometrial cancer and undergoing investigations and treatment were collected to test the software with their cancer diagnosis not revealed to the software developers.
The forest plots allowed an overall relative risk and percentage risk to be calculated from all the risk data gathered from the studies. A neural network computational model to determine percentage risk for individual patients was developed, implemented, and evaluated.
The results show that the greatest percentage increased risk was due to BMI being above 25, with the risk increasing as BMI increases. A BMI of 25 or over gave an increased risk of 2.01%, a BMI of 30 or over gave an increase of 5.24%, and a BMI of 40 or over led to an increase of 6.9%. PCOS was the second highest increased risk at 4.2%. Diabetes, which is incidentally also linked to an increased BMI, gave a significant increased risk along with null parity and noncontinuous HRT of 1.54%, 1.2%, and 0.56% respectively. Decreased risk due to contraception was greatest with IUD (intrauterine device) and IUPD (intrauterine progesterone device) at -1.34% compared to -0.9% with oral. Continuous HRT at -0.75% and parity at -0.9% also decreased the risk. Using open-source patient data to test our computational model to determine risk, our results showed that the model is 98.6% accurate with an algorithm sensitivity 75% on average.
In this study, we successfully determined the rank order of risk factors for endometrial cancer and calculated a pooled risk and risk percentage for each factor using a statistical meta-analysis approach. Then, using a computer neural network model system, we were able to model the overall increase or decreased risk of cancer and predict the cancer diagnosis for particular patients to an accuracy of over 98%. The neural network model developed in this study was shown to be a potentially useful tool in determining the percentage risk and predicting the possibility of a given patient developing endometrial cancer. As such, it could be a useful tool for clinicians to use in conjunction with other biomarkers in determining which patients warrant further preventative interventions to avert progressing to endometrial cancer. This result would allow for a reduction in the number of unnecessary invasive tests on patients. The model may also be used to suggest interventions to decrease the risk for a particular patient. The sensitivity of the model limits it at this stage due to the small percentage of positive cases in the datasets; however, since this model utilizes a neural network machine learning algorithm, it can be further improved by providing the system with more and larger datasets to allow further refinement of the neural network.
在本研究中，我们希望确定子宫内膜癌风险因素的排序，并使用统计荟萃分析方法计算每个因素的合并风险和风险百分比。下一步是设计一个神经网络计算机模型，以预测个体患者患癌风险的总体增加或降低情况。这将有助于确定该预测是否可用作决定患者是否应考虑进行检测以及预测诊断的工具，并向患者建议预防措施。
对现有数据进行荟萃分析以计算相对风险，随后基于神经网络算法设计并实施风险预测计算模型。
荟萃分析数据来自世界各地的不同背景。用于测试模型的原始数据是从医院诊所收集的。
收集了40例目前疑似患有子宫内膜癌并正在接受检查和治疗的患者病历数据，在向软件开发人员隐瞒癌症诊断的情况下对软件进行测试。
森林图允许根据从研究中收集的所有风险数据计算总体相对风险和风险百分比。开发、实施并评估了一个用于确定个体患者风险百分比的神经网络计算模型。
结果表明，风险增加百分比最大的是BMI高于25，且风险随BMI升高而增加。BMI为25或以上时，风险增加2.01%；BMI为30或以上时，风险增加5.24%；BMI为40或以上时，风险增加6.9%。多囊卵巢综合征（PCOS）是风险增加第二高的因素，为4.2%。糖尿病（其也与BMI升高相关）、未生育以及非连续激素替代疗法（HRT）导致的风险显著增加，分别为1.54%、1.2%和0.56%。避孕措施中，宫内节育器（IUD）和宫内孕激素装置（IUPD）导致的风险降低最大，为-1.34%，而口服避孕药为-0.9%。连续HRT为-0.75%，生育为-0.9%也降低了风险。使用开源患者数据测试我们的计算模型以确定风险，结果表明该模型的准确率为98.6%，算法平均灵敏度为75%。
在本研究中，我们成功确定了子宫内膜癌风险因素的排序，并使用统计荟萃分析方法计算了每个因素的合并风险和风险百分比。然后，使用计算机神经网络模型系统，我们能够模拟患癌风险的总体增加或降低情况，并以超过98%的准确率预测特定患者的癌症诊断。本研究中开发的神经网络模型被证明是确定风险百分比和预测特定患者患子宫内膜癌可能性的潜在有用工具。因此，它可能是临床医生在结合其他生物标志物确定哪些患者需要进一步预防性干预以避免发展为子宫内膜癌时的有用工具。这一结果将减少对患者进行不必要的侵入性检查的数量。该模型还可用于建议降低特定患者风险的干预措施。由于数据集中阳性病例的百分比很小，该模型在现阶段的灵敏度限制了它；然而，由于该模型使用神经网络机器学习算法，可以通过为系统提供更多更大的数据集来进一步改进，以进一步优化神经网络。
