Risk Identification Models Of Water Intoxication During Hysteroscopic Adhesiolysis Based On Multimodal Data

Background: Water intoxication is a potentially fatal hysteroscopic complication due to excessive fluid absorption.Timely risk identification of water intoxication is the key point and difficulty of complication prevention.Fluid absorption during hysteroscopy is affected by many factors.The faster the real-time fluid absorption rate during hysteroscopy,the higher the risk of water intoxication during the subsequent operation.However,due to the operator’s lack of mastery of the real-time absorption rate,the rate’s change with the operation,the status of each factor,and the magnitude of its influence,it is difficult to identify the risk in time during hysteroscopy,making it impossible to intervene in time to slow fluid absorption,preventing water intoxication.Hysteroscopic adhesiolysis(HA)is the preferred treatment for intrauterine adhesions(IUAs).However,when dealing with moderate to severe IUAs,water intoxication is prone to occur during HA,especially when most of the uterine cavity is closed due to adhesions or the fallopian tube ostium is difficult to see.It is difficult to identify the original uterine cavity layer,and the surgical wound is large.Therefore,this study intended to explore how to assist surgeons in timely identifying the risk of water intoxication during HA.Objective: To develop a device that can monitor the fluid absorption during hysteroscopic surgery in real time,including the fluid deficit and absorption rate,and establish a multi-modal dataset that includes clinical data of IUAs patients,fluid absorption data during HA surgery,and surgical videos;and on this basis,to develop and validate risk identification models of water intoxication during HA from three dimensions: changes detection of fluid absorption rate,factors analysis of fluid absorption,and hysteroscopic images identification of surgical injury,so as to assist surgeons in timely risk identification and intervention,effectively preventing water intoxication.Methods:(1)Device development and dataset establishment: The device was designed based on the fluid volume balance method,and was divided into four modules: fluid collection,fluid measurement,data processing,and information display.They were used for the outflow collection,the measurement of inflow and outflow,calculation and data storage of fluid deficit and absorption rate,display of fluid absorption(including real-time values and changing trends of fluid deficit and absorption rate),and voice alarm for abnormal values,respectively.The standard weights(2000g and 5000g)were measured by 20 times using the inflow and outflow load cells in the fluid measurement module,and the difference between the measured and actual value was compared to verify the measurement accuracy of the device.Subsequently,patients with IUAs who visited the Third Xiangya Hospital of Central South University between May 2020 and May 2021 were prospectively included.The fluid absorption during HA was monitored using the device.At the same time,the operation process under the hysteroscopy was filmed,and the patient’s clinical data(including basic information,gynecological diseases,reproductive history and surgical data)were collected to establish a multimodal dataset,providing a data basis for subsequent model research.(2)Changes detection of the fluid absorption rate: for the HA surgeries where the initial value of perfusion pressure is 100 mm Hg,the time series change point detection algorithm based on a sliding window was used to establish a detection model for the changes of fluid absorption status to monitor the trend of the absorption rate with the operation time and detect the change point of the fluid absorption status with a significant increase in the risk coefficient of water intoxication.Subsequently,the precision and recall of the change points detection were calculated to evaluate the model performance.Combined with the operation videos,the factors leading to the increased risk coefficient of water intoxication during HA were analyzed.(3)Factors analysis of the fluid absorption: four machine learning algorithms(ridge logistic regression,decision tree,gradient boosting decision tree and random forest)were used to establish a risk identification model for the fluid absorption status,judging the water intoxication risk level of the fluid status(low risk or high risk)during HA based on 12 clinical features(age,gravidity,mean arterial pressure,preoperative misoprostol use,mean arterial pressure,stenosis of lower uterine cavity,visibility of preoperative fallopian tube ostium,density of endometrial glandular openings,uterine cavity depth,uterine cavity width,severity of IUAs,and myometrium damage).Subsequently,the area under the receiver operating curve(AUC)was calculated to evaluate the model’s identification performance.And the SHAP(Shapley Additive ex Planations)method was used to analyze the importance of each clinical feature in the model with the highest mean AUC.(4)Hysteroscopic images recognition of surgical injury: the deep learning convolutional neural network was used to establish a image recognition model of myometrium damage to judge whether patients have myometrium damage,a dangerous clinical feature,based on hysteroscopic images during HA.Subsequently,the precision,recall,accuracy,and AUC were calculated to evaluate the model’s recognition performance.Results:(1)Device development and dataset establishment: There was no significant difference between the actual values of the standard weights and the measured values of the inflow and outflow load cells in the fluid measurement module of the hysteroscopic distension fluid absorption monitoring device(P > 0.05).A total of 525 IUAs patients were prospectively included,and 61 of them triggered an abnormal alarm of the device during HA(fluid absorption rate ≥ 80 m L/min).Statistical results of the HA multimodal dataset: mild,moderate,and severe IUAs were 187,218,and 120 cases,respectively;the average age of the patients was 32.33 ± 4.63 years old;the average fluid deficit was 113.50(52.89,289.40)m L;The operation duration is 622(410,859)s.(2)Changes detection of the fluid absorption rate: 244 patients with moderate to severe IUAs were included from the multimodal dataset.In the training and test set,the precision of fluid absorption status’ s change points detecting by the model was 93.30% and 87.88%,and the recall rate was 95.10% and 90.63%,respectively.The increased risk coefficient of water intoxication in patients during HA mainly occurred in three surgical scenarios: when separating adhesions in the nearly closed or completely closed lower segment of the uterine cavity,when separating adhesions at the closed uterine horns,and when separating peripheral adhesions and loosening the narrow ring of scar.Myometrium damage was a common factor in the three surgical scenarios.The other factors were successful access to the upper segment of the uterine cavity,successful exposure of the fallopian tube ostium,higher intrauterine pressure due to increased fluid perfusion pressure,clamping of the external os of the cervix,or obvious stenosis of the lower segment of the uterine cavity.(3)Factors analysis of the fluid absorption: 308 patients with moderate to severe IUAs were included from the multimodal dataset.Based on the 12 clinical features,the random forest model had the best performance in identifying the water intoxication risk level of the patients’ fluid absorption status,with an average AUC of 0.82±0.07.Myometrium damage,uterine cavity width,and lower uterine cavity stenosis are the three most important clinical features for the water intoxication risk level.In addition,myometrium damage,the width of the uterine cavity adapted to large and larger intrauterine devices,and the closed lower segment of the uterine cavity all increase the possibility of patients in a high-risk water intoxication status.(4)Hysteroscopic images recognition of surgical injury: 296 patients with moderate to severe IUAs were included in the multimodal dataset,and 1481 hysteroscopic images were intercepted,including 706 images of myometrium damage,and 775 normal images without myometrium damage.In the training set and test set,the precision of the image recognition model is 94.02% and 91.94%,the recall rate is 91.46% and90.96%,the accuracy is 93.30% and 91.42%,and the AUC value is 0.989.and 0.971,respectively.Conclusions:(1)This study successfully developed a hysteroscopic fluid absorption monitoring device with real-time monitoring,trend display,abnormal alarm and data storage functions,and successfully established a multi-dimensional risk identification model of water intoxication during HA based on multi-modal data.(2)The model can assist the surgeon in timely identifying the risk of water intoxication in IUAs patients during HA from three dimensions,so as to intervene and prevent water intoxication in a timely manner:(1)timely and accurately detect that the water intoxication risk coefficient of the patient’s fluid absorption status has increased significantly based on the data of fluid absorption rate;(2)Based on 12 clinical features,accurately identify that patients’ fluid absorption status were at high risk of water intoxication,and analyze the main influencing factors;(3)Based on the hysteroscopic image,accurately determine that the patient has myometrium damage,a risk factor of water intoxication.