Dose volume histogram prediction method for organ at risk in VMAT planning of nasopharyngeal carcinoma based on equivalent uniform dose / 中华放射肿瘤学杂志

Objective:To evaluate the practicability of dose volume histogram (DVH) prediction model for organ at risk (OAR) of radiotherapy plan by minimizing the cost function based on equivalent uniform dose (EUD).Methods:A total of 66 nasopharyngeal carcinoma (NPC) patients received volume rotational intensity modulated arc therapy (VMAT) at Cancer Hospital & Shenzhen Hospital, Chinese Academy of Medical Sciences from 2020 to 2021 were retrospectively selected for this study. Among them, 50 patients were used to train the recurrent neutral network (RNN) model and the remaining 16 cases were used to test the model. DVH prediction model was constructed based on RNN. A three-dimensional equal-weighted 9-field conformal plan was designed for each patient. For each OAR, the DVHs of individual fields were acquired as the model input, and the DVH of VMAT plan was regarded as the expected output. The prediction error obtained by minimizing EUD-based cost function was employed to train the model. The prediction accuracy was characterized by the mean and standard deviation between predicted and true values. The plan was re-optimized for the test cases based on the DVH prediction results, and the consistency and variability of the EUD and DVH parameters of interest (e.g., maximum dose for serial organs such as the spinal cord) were compared between the re-optimized plan and the original plan of OAR by the Wilcoxon paired test and box line plots.Results:The neural network obtained by training the cost function based on EUD was able to obtain better DVH prediction results. The new plan guided by the predicted DVH was in good agreement with the original plan: in most cases, the D 98% in the planning target volume (PTV) was greater than 95% of the prescribed dose for both plans, and there was no significant difference in the maximum dose and EUD in the brainstem, spinal cord and lens (all P>0.05). Compared with the original plan, the average reduction of optic chiasm, optic nerves and eyes in the new plans reached more than 1.56 Gy for the maximum doses and more than 1.22 Gy for EUD, and the average increment of temporal lobes reached 0.60 Gy for the maximum dose and 0.30 Gy for EUD. Conclusion:The EUD-based loss function improves the accuracy of DVH prediction, ensuring appropriate dose targets for treatment plan optimization and better consistency in the plan quality.

Recurrent neural network prediction on clinical usage of red blood cells / 中国输血杂志

【Objective】 To explore the prediction of clinical red blood cells (RBCs) consumption under major public health emergencies, so as to provide scientific basis for blood collection and blood inventory management. 【Methods】 The clinical consumption of different types of RBCs in Yichang from 2001 to 2017 was analyzed and modeled using the recurrent neural network (RNN) model, and the clinical RBCs consumption between January 2019 and December 2021(36 months) were scientifically predicted. 【Results】 The RNN model showed good prediction performance. The root mean square errors (RMSE) of RNN prediction values of A, B, O, AB type of RBCs were 156.7, 133.4, 204.2 and 51.3, respectively, with the average relative errors (MRE) at 6.4%, 6.6%, 8.5% and 7.1%, respectively. The model predicted the changing trend of RBCs consumption during the first round of COVID-19 outbreak (January to June, 2020) and forecasted the lowest level of consumption in February 2020 and a subsequent recovery in growth. The prediction of RBCs consumption during the second round of COVID-19 pandemic (January to June, 2021) was of high accuracy. For example, the relative errors of RNN models for A type RBCs consumption were 5.2% in Feb 2021 (the lowest level, 1 621.5 U) and 2.5% in May 2021 (the highest level, 2 397.0 U). 【Conclusion】 The artificial intelligence RNN model can predict clinical RBCs consumption well under major public health emergencies.

Knee Joint Moment Estimation During Walking via Wearable Inertial Sensor Network / 医用生物力学

Objective To estimate knee adduction moment (KAM) and knee flexion moment (KFM) under different gait test conditions via an inertial sensor network (ISN). Methods Twelve healthy young male subjects wore eight inertial sensors (located in the trunk, pelvis, both thighs, both shanks, both feet) and walked under different test conditions (changing foot progression angle, trunk sway angle, step width and walking speed). An ISN was used to extract biomechanical features as the input of recurrent neural network (RNN), so as to estimate the KAM and KFM. Results The overall KAM estimation accuracy: relative root mean square error (rRMSE) was 8.54% and r=0.84. The overall KFM estimation accuracy was rRMSE=6.40% and r=0.94. Conclusions The model can be used as the basis for load estimation of knee joints out of the lab and its potential application includes gait training and rehabilitation assessment after knee surgery.

Research on prediction of daily admissions of respiratory diseases with comorbid diabetes in Beijing based on long short-term memory recurrent neural network / 浙江大学学报·医学版

To compare the performance of generalized additive model (GAM) and long short-term memory recurrent neural network (LSTM-RNN) on the prediction of daily admissions of respiratory diseases with comorbid diabetes. Daily data on air pollutants, meteorological factors and hospital admissions for respiratory diseases from Jan 1st, 2014 to Dec 31st, 2019 in Beijing were collected. LSTM-RNN was used to predict the daily admissions of respiratory diseases with comorbid diabetes, and the results were compared with those of GAM. The evaluation indexes were calculated by five-fold cross validation. Compared with the GAM, the prediction errors of LSTM-RNN were significantly lower [root mean squared error (RMSE): 21.21±3.30 vs. 46.13±7.60, <0.01; mean absolute error (MAE): 14.64±1.99 vs. 36.08±6.20, <0.01], and the value was significantly higher (0.79±0.06 vs. 0.57±0.12, <0.01). In gender stratification, RMSE, MAE and values of LSTM-RNN were better than those of GAM in predicting female admission (all <0.05), but there were no significant difference in predicting male admission between two models (all >0.05). In seasonal stratification, RMSE and MAE of LSTM-RNN were lower than those of GAM in predicting warm season admission (all <0.05), but there was no significant difference in value (>0.05). There were no significant difference in RMSE, MAE and between the two models in predicting cold season admission (all >0.05). In the stratification of functional areas, the RMSE, MAE and values of LSTM-RNN were better than those of GAM in predicting core area admission (all <0.05). has lower prediction errors and better fitting than the GAM, which can provide scientific basis for precise allocation of medical resources in polluted weather in advance.

A hybrid attention temporal sequential network for sleep stage classification / 生物医学工程学杂志

Sleep stage classification is a necessary fundamental method for the diagnosis of sleep diseases, which has attracted extensive attention in recent years. Traditional methods for sleep stage classification, such as manual marking methods and machine learning algorithms, have the limitations of low efficiency and defective generalization. Recently, deep neural networks have shown improved results by the capability of learning complex pattern in the sleep data. However, these models ignore the intra-temporal sequential information and the correlation among all channels in each segment of the sleep data. To solve these problems, a hybrid attention temporal sequential network model is proposed in this paper, choosing recurrent neural network to replace traditional convolutional neural network, and extracting temporal features of polysomnography from the perspective of time. Furthermore, intra-temporal attention mechanism and channel attention mechanism are adopted to achieve the fusion of the intra-temporal representation and the fusion of channel-correlated representation. And then, based on recurrent neural network and inter-temporal attention mechanism, this model further realized the fusion of inter-temporal contextual representation. Finally, the end-to-end automatic sleep stage classification is accomplished according to the above hybrid representation. This paper evaluates the proposed model based on two public benchmark sleep datasets downloaded from open-source website, which include a number of polysomnography. Experimental results show that the proposed model could achieve better performance compared with ten state-of-the-art baselines. The overall accuracy of sleep stage classification could reach 0.801, 0.801 and 0.717, respectively. Meanwhile, the macro average F1-scores of the proposed model could reach 0.752, 0.728 and 0.700. All experimental results could demonstrate the effectiveness of the proposed model.

An Evaluation of Lower Limb Motor Ability Based on Xception-LSTM / 中国康复理论与实践

Objective:To establish an algorithm to quantitatively evaluate the lower limb motor ability. Methods:From August, 2016 to March, 2017, 40 subjects were divided into young healthy group (n = 20), middle-aged group (n = 10) and elderly group (n = 10). The gait video, knee angle and ground reaction force of the subjects were collected, and the gait contour was extracted from the gait video by using the ViBe algorithm. The gait image feature was extracted by Xception-LSTM, and fused it with the knee joint angle and the ground reaction force in the feature layer. The fusion features were reduced in dimension by kernel principal component analysis, and the gait ability score (GAS) was established. All the subjects were assessed with Wisconsin Gait Scale (WGS). Results:GAS was less in middle-aged group and elderly group than in the young healthy group (t > 4.164, P < 0.01), and was less in the elderly group than in the middle-aged group (t = 7.338, P < 0.01). GAS was negative correlated with the score of WGS (r = -0.91, P < 0.01). Conclusion:The lower limb exercise ability could be quantified with GAS, which may be applied in developing rehabilitation and fitting walking aids.

Personalized Urination Activity Recognition Based on a Recurrent Neural Network Using Smart Band / 대한배뇨장애요실금학회지

PURPOSE: Though it is very important obtaining exact data about patients’ voiding patterns for managing voiding dysfunction, actual practice is very difficult and cumbersome. In this study, data about urination time and interval measured by smart band device on patients’ wrist were collected and analyzed to resolve the clinical arguments about the efficacy of voiding diary. By developing a smart band based algorithm for recognition of complex and serial pattern of motion, this study aimed to explore the feasibility of measurement the urination time and intervals for voiding dysfunction management. METHODS: We designed a device capable of recognizing urination time and intervals based on specific postures of the patient and consistent changes in posture. These motion data were obtained by a smart band worn on the wrist. An algorithm that recognizes the repetitive and common 3-step behavior for urination (forward movement, urination, backward movement) was devised based on the movement and tilt angle data collected from a 3-axis accelerometer. The sequence of body movements during voiding has consistent temporal characteristics, so we used a recurrent neural network and long short-term memory based framework to analyze the sequential data and to recognize urination time. Real-time data were acquired from the smart band, and for data corresponding to a certain duration, the value of the signals was calculated and then compared with the set analysis model to calculate the time of urination. A comparative study was conducted between real voiding and device-detected voiding to assess the performance of the proposed recognition technology. RESULTS: The accuracy of the algorithm was calculated based on clinical guidelines established by urologists. The accuracy of this detecting device was high (up to 94.2%), proving the robustness of the proposed algorithm. CONCLUSIONS: This urination behavior recognition technology showed high accuracy and could be applied in clinical settings to characterize patients’ voiding patterns. As wearable devices are developed and generalized, algorithms detecting consistent sequential body movement patterns reflecting specific physiologic behavior might be a new methodology for studying human physiologic behavior.

Integration of A Deep Learning Classifier with A Random Forest Approach for Predicting Malonylation Sites / 基因组蛋白质组与生物信息学报·英文版

As a newly-identified protein post-translational modification, malonylation is involved in a variety of biological functions. Recognizing malonylation sites in substrates represents an initial but crucial step in elucidating the molecular mechanisms underlying protein malonylation. In this study, we constructed a deep learning (DL) network classifier based on long short-term memory (LSTM) with word embedding (LSTM) for the prediction of mammalian malonylation sites. LSTM performs better than traditional classifiers developed with common pre-defined feature encodings or a DL classifier based on LSTM with a one-hot vector. The performance of LSTM is sensitive to the size of the training set, but this limitation can be overcome by integration with a traditional machine learning (ML) classifier. Accordingly, an integrated approach called LEMP was developed, which includes LSTM and the random forest classifier with a novel encoding of enhanced amino acid content. LEMP performs not only better than the individual classifiers but also superior to the currently-available malonylation predictors. Additionally, it demonstrates a promising performance with a low false positive rate, which is highly useful in the prediction application. Overall, LEMP is a useful tool for easily identifying malonylation sites with high confidence. LEMP is available at http://www.bioinfogo.org/lemp.

Deep Learning in Medical Imaging: General Overview

The artificial neural network (ANN)–a machine learning technique inspired by the human neuronal synapse system–was introduced in the 1950s. However, the ANN was previously limited in its ability to solve actual problems, due to the vanishing gradient and overfitting problems with training of deep architecture, lack of computing power, and primarily the absence of sufficient data to train the computer system. Interest in this concept has lately resurfaced, due to the availability of big data, enhanced computing power with the current graphics processing units, and novel algorithms to train the deep neural network. Recent studies on this technology suggest its potentially to perform better than humans in some visual and auditory recognition tasks, which may portend its applications in medicine and healthcare, especially in medical imaging, in the foreseeable future. This review article offers perspectives on the history, development, and applications of deep learning technology, particularly regarding its applications in medical imaging.