Median RMSE Research Articles

Therapeutic Drug Monitoring, Population Pharmacokinetics Models, and External Validation of High-Dose Methotrexate in Pediatric Acute Lymphoblastic Leukemia.

High-dose methotrexate (MTX) is used to treat pediatric acute lymphoblastic leukemia (ALL). The drug has a low therapeutic index and a highly interindividual variability in systemic exposure. These characteristics necessitate dose adjustments and therapeutic drug monitoring protocols, while population pharmacokinetic (POP/PK) models may enable more precise drug dosing. Therefore, we assessed the performance of external POP/PK models in ALL children receiving high-dose MTX. We retrospectively harvested clinical and laboratory data from ALL children during their first two cycles of chemotherapy. A POP/PK model was elaborated using the Monolix suite 2024R1. External models were selected from PUBMED based on strict inclusion/exclusion criteria, and their fit to the actual data was assessed by calculating bias (percentage prediction error [PE%]) and precision (percentage root mean squared error [RMSE%]). Thirty-seven ALL children participated in the study (18 males, median age 5.1 years, range 1.7-15.2 years), and six external POP/PK models were chosen. Except for one model (median PE% value, -97.45%), all models exhibited acceptable bias (median PE% values, -4.17%-2.67%), despite none of them demonstrating good precision (median RMSE% values, 89.19%-120.40%). External models should be accurately evaluated before they are implemented in clinical practice, even when patients share very similar characteristics.

Accuracy Evaluation of 3D Pose Reconstruction Algorithms Through Stereo Camera Information Fusion for Physical Exercises with MediaPipe Pose.

In recent years, significant research has been conducted on video-based human pose estimation (HPE). While monocular two-dimensional (2D) HPE has been shown to achieve high performance, monocular three-dimensional (3D) HPE poses a more challenging problem. However, since human motion happens in a 3D space, 3D HPE offers a more accurate representation of the human, granting increased usability for complex tasks like analysis of physical exercise. We propose a method based on MediaPipe Pose, 2D HPE on stereo cameras and a fusion algorithm without prior stereo calibration to reconstruct 3D poses, combining the advantages of high accuracy in 2D HPE with the increased usability of 3D coordinates. We evaluate this method on a self-recorded database focused on physical exercise to research what accuracy can be achieved and whether this accuracy is sufficient to recognize errors in exercise performance. We find that our method achieves significantly improved performance compared to monocular 3D HPE (median RMSE of 30.1 compared to 56.3, p-value below 10-6) and can show that the performance is sufficient for error recognition.

Multi-Step Ahead Water Level Forecasting Using Deep Neural Networks

Stream gauge height (water level) is a significant indicator for forecasting future floods. Flooding occurs when the water level exceeds the flood stage. Predicting imminent floods can save lives, protect infrastructure, and improve road traffic management and transportation. Deep neural networks have been increasingly used in this domain due to their predictive capabilities in capturing complex features and interdependencies. This study employs four distinct models—Multi-Layer Perceptron (MLP), Long Short-Term Memory (LSTM), transformer, and LSTNet—with MLP serving as the baseline model to forecast water levels. The models are trained using data from 20 distinct river gages across the state of Missouri to ensure consistent performance. Random search optimization is employed for hyperparameter tuning. The prediction intervals are set at 4, 6, 8, and 10 (each interval equivalent to 30 min) to ensure that performance results are robust and not due to random weight initialization or suboptimal hyperparameters and are consistent throughout different prediction intervals. The findings of this study indicate that the LSTNet model leads to a better performance than the other models, with a median RMSE of 0.00724, 0.00959, 0.01204, and 0.01230 for the 4, 6, 8, and 10 intervals, respectively. As climate change leads to localized storms driven by atmospheric shifts, water level fluctuations are becoming increasingly extreme, further exacerbating data drift in real-world datasets. The LSTNet model demonstrates superior performance in terms of RMSE, MAE, and the correlation coefficient across all prediction intervals when forecasting water levels under data drift conditions.

Evaluation of global seamless soil moisture products over China: A perspective of soil moisture sensitivity to precipitation

Accurate and timely information about soil moisture (SM) is not only critical to the study of climate change, but is also significant for agricultural production, drought monitoring, and hydrometeorological predictions. With the development of satellite remote sensing techniques, land surface models, and data assimilation techniques, a number of global seamless soil moisture products have been released with highly variable accuracy. A comprehensive evaluation of these products is necessary to select a suitable SM product for use. In this study, we provided such an evaluation of four global seamless SM products (SMAP-L4, ERA5-Land, GLDAS-Noah, and GLEAM) over China based on the stations from the automatic soil moisture observation (ASMO) system. Compared to previous studies that focused on a direct comparison between SM products and in situ SM observations, we specifically evaluated these products from the perspective of SM sensitivity to precipitation. This is because in situ precipitation observations are more spatially representative than in situ SM observations. Precipitation is a regional event, therefore in situ precipitation observations can be better scale-matched with coarse resolution SM products compared to in situ SM observations. Further, in addition to surface soil moisture (SSM), root zone soil moisture (RZSM) was evaluated in this study. The combination of all statistics illustrates that SMAP-L4 and ERA5-Land generally performed better than GLDAS-Noah and GLEAM. Regardless of the SSM or RZSM estimates, SMAP-L4 had the lowest median bias and RMSE of the four products, and ERA5-Land had the lowest median correlation coefficient (R) value. The lowest median ubRMSE was obtained by SMAP-L4 in SSM evaluation and ERA5-Land performed better than other RZSM products with the lowest median ubRMSE. From the perspective of SM sensitivity to precipitation, SMAP-L4 can capture daily precipitation dynamics most effectively, closely followed by ERA5-Land. This partially explains the superiority of SMAP-L4 and ERA5-Land over GLDAS-Noah and GLEAM for SM monitoring. The results of the metrics under different precipitation conditions varied considerably among the products. But while considering of ubRMSE alone, all products performed better under arid and humid conditions than to semi-arid and semi-humid conditions. This may be due to the fact that SM products had a limited ability to capture the SM dynamics under moderate precipitation. These findings may provide new insights into the selection of appropriate SM products and SM estimation in the future.

A machine learning model for estimating the temperature of small rivers using satellite-based spatial data

The influence of anthropogenic activity and land cover alteration on stream temperatures has major ecological implications, such as limiting fish survival. While these ecological impacts have been extensively studied at varying spatial scales for major rivers, our understanding of the range and complexity of this relationship across climates and geographies for smaller rivers (less than ∼60 m wide) remains limited. This is in part because although such rivers comprise a vast majority (∼97%) of total stream length, most smaller rivers lack routine temperature observations despite extensive gage networks. While existing gage networks in many regions are sufficient to support satellite remote sensing-based modeling of small stream temperatures, most large regions including the contiguous United States do not have modeled data products at high spatial and temporal resolution for small rivers.In this study, we developed a statistical model, TempEst (“temperature estimation”), to estimate monthly average water temperatures for smaller rivers and evaluated the model with rivers ranging from 3 m to 1000 m (1 km) in width, using land cover, topography, and satellite-based land surface temperature. The rivers chosen for development of the model are representative of a range of urban and rural land uses and cover eight US Environmental Protection Agency Level I ecoregions. The model was calibrated and validated using observed stream temperatures from the United States Geological Survey. TempEst is able to predict monthly average temperature for streams of any size with an overall median RMSE of about 1.5 °C and bias of 0%. Model accuracy has a consistent trend with training gauge network density, with median RMSE increasing to about 2.0 °C in more sparsely-gauged regions. As a simple demonstration, we used TempEst to estimate thermal suitability conditions for cutthroat trout in Rocky Mountain streams in the United States. The developed model, available as open-source R (model) and Google Earth Engine (data retrieval) scripts, will facilitate the study of stream temperature behaviors at higher resolutions than previously available across the contiguous United States and can be easily adapted to support global river systems.

Predicting Tibia-Fibula Geometry and Density From Anatomical Landmarks Via Statistical Appearance Model: Influence of Errors on Finite Element-Calculated Bone Strain.

State-of-the-art participant-specific finite element models require advanced medical imaging to quantify bone geometry and density distribution; access to and cost of imaging is prohibitive to the use of this approach. Statistical appearance models may enable estimation of participants' geometry and density in the absence of medical imaging. The purpose of this study was to: (1) quantify errors associated with predicting tibia-fibula geometry and density distribution from skin-mounted landmarks using a statistical appearance model and (2) quantify how those errors propagate to finite element-calculated bone strain. Participant-informed models of the tibia and fibula were generated for thirty participants from height and sex and from twelve skin-mounted landmarks using a statistical appearance model. Participant-specific running loads, calculated using gait data and a musculoskeletal model, were applied to participant-informed and CT-based models to predict bone strain using the finite element method. Participant-informed meshes illustrated median geometry and density distribution errors of 4.39-5.17 mm and 0.116-0.142 g/cm3, respectively, resulting in large errors in strain distribution (median RMSE = 476-492 με), peak strain (limits of agreement =±27-34%), and strained volume (limits of agreement =±104-202%). These findings indicate that neither skin-mounted landmark nor height and sex-based predictions could adequately approximate CT-derived participant-specific geometry, density distribution, or finite element-predicted bone strain and therefore should not be used for analyses comparing between groups or individuals.

Where to mount the IMU? Validation of joint angle kinematics and sensor selection for activities of daily living

We validate the OpenSense framework for IMU-based joint angle estimation and furthermore analyze the framework's ability for sensor selection and optimal positioning during activities of daily living (ADL). Personalized musculoskeletal models were created from anthropometric data of 19 participants. Quaternion coordinates were derived from measured IMU data and served as input to the simulation framework. Six ADLs, involving upper and lower limbs were measured and a total of 26 angles analyzed. We compared the joint kinematics of IMU-based simulations with those of optical marker-based simulations for most important angles per ADL. Additionally, we analyze the influence of sensor count on estimation performance and deviations between joint angles, and derive the best sensor combinations. We report differences in functional range of motion (fRoMD) estimation performance. Results for IMU-based simulations showed MAD, RMSE, and fRoMD of 4.8°, 6.6°, 7.2° for lower limbs and for lower limbs and 9.2°, 11.4°, 13.8° for upper limbs depending on the ADL. Overall, sagittal plane movements (flexion/extension) showed lower median MAD, RMSE, and fRoMD compared to transversal and frontal plane movements (rotations, adduction/abduction). Analysis of sensor selection showed that after three sensors for the lower limbs and four sensors for the complex shoulder joint, the estimation error decreased only marginally. Global optimum (lowest RMSE) was obtained for five to eight sensors depending on the joint angle across all ADLs. The sensor combinations with the minimum count were a subset of the most frequent sensor combinations within a narrowed search space of the 5% lowest error range across all ADLs and participants. Smallest errors were on average < 2° over all joint angles. Our results showed that the open-source OpenSense framework not only serves as a valid tool for realistic representation of joint kinematics and fRoM, but also yields valid results for IMU sensor selection for a comprehensive set of ADLs involving upper and lower limbs. The results can help researchers to determine appropriate sensor positions and sensor configurations without the need for detailed biomechanical knowledge.

A modified Michaelis-Menten equation estimates growth from birth to 3 years in healthy babies in the USA

BackgroundStandard pediatric growth curves cannot be used to impute missing height or weight measurements in individual children. The Michaelis–Menten equation, used for characterizing substrate-enzyme saturation curves, has been shown to model growth in many organisms including nonhuman vertebrates. We investigated whether this equation could be used to interpolate missing growth data in children in the first three years of life and compared this interpolation to several common interpolation methods and pediatric growth models.MethodsWe developed a modified Michaelis–Menten equation and compared expected to actual growth, first in a local birth cohort (N = 97) then in a large, outpatient, pediatric sample (N = 14,695).ResultsThe modified Michaelis–Menten equation showed excellent fit for both infant weight (median RMSE: boys: 0.22 kg [IQR:0.19; 90% < 0.43]; girls: 0.20 kg [IQR:0.17; 90% < 0.39]) and height (median RMSE: boys: 0.93 cm [IQR:0.53; 90% < 1.0]; girls: 0.91 cm [IQR:0.50;90% < 1.0]). Growth data were modeled accurately with as few as four values from routine well-baby visits in year 1 and seven values in years 1–3; birth weight or length was essential for best fit. Interpolation with this equation had comparable (for weight) or lower (for height) mean RMSE compared to the best performing alternative models.ConclusionsA modified Michaelis–Menten equation accurately describes growth in healthy babies aged 0–36 months, allowing interpolation of missing weight and height values in individual longitudinal measurement series. The growth pattern in healthy babies in resource-rich environments mirrors an enzymatic saturation curve.

Modeling Patient-specific Apnea-bradycardia Patterns in Preterm Newborn.

Preterm infants are particularly exposed to severe cardio-respiratory events, associating apnea with bradycardia and oxygen desaturation. A patient-specific and event-specific model-based approach is proposed in this work to analyze the acute heart rate response to apnea-bradycardia events in preterm newborn. A novel model integrating the main cardio-respiratory interactions which are specific to the neonatal period is proposed. An evolutionary algorithm is applied to estimate patient-specific model parameters from a database of 37 apnea-bradycardia episodes acquired from 10 preterm newborns. Unsupervised clustering (K-means) was applied to the identified parameters to define phenogroups of cardio-respiratory responses to apnea. A significant correspondence was observed between simulated and experimental heart rate series in all identifications (median RMSE = 8.85 bpm). Three clusters of parameters were found and were associated to three different pathophysiological dynamics related to apnea-bradycardia. The proposed method, based on patient and event-specific model parameter identification, provides a novel approach to characterize bradycardia dynamics in response to apnea, opening the way to the proposal of new personalized diagnosis and treatment possibilities in this particularly sensitive population.

Evaluation of a regional crop model implementation for sub-national yield assessments in Kenya

CONTEXTCropping system models can be used to both assess regional food security and to monitor and predict agricultural drought. Agriculture in Kenya is extremely important to both the economy and food security of the country. OBJECTIVEThis study evaluated a regional implementation of a widely used crop model, the Decision Support System for Agrotechnology Transfer (DSSAT), within a coupled modeling framework, the Regional Hydrologic Extremes Assessment System (RHEAS), over Kenya. The goal of this study was to assess the ability of RHEAS to simulate the annual variability of maize yields at the county level and evaluate the uncertainty inherent in the model and inputs. METHODSThe RHEAS system implements a stochastic ensemble approach to account for field scale variabilities in crop management practices and underlying soil and weather conditions. Satellite-derived datasets were used to evaluate the land surface component of the system and seasonally disaggregated yield for 5 years was used to assess the performance of the cropping system model. RESULTS AND CONCLUSIONSThe median correlation between RHEAS and satellite-derived soil moisture and evapotranspiration estimates were 0.78, and 0.51, respectively, indicating that the model is able to capture the key drivers of the hydrological budget. Overall, RHEAS simulated yearly yield variations with a median correlation of 0.7 with reported yields, with the best performance in the short rains season. However, across both seasons, the RHEAS model was positively biased on the order of ∼1.6 MT/ha. The overall median unbiased RMSE was 0.66 MT/ha. The RHEAS system shows skill at simulating extreme departures in anomalies, and a majority of the time (62.5%) the reported yields fall within the interquartile range of the simulations. SIGNIFICANCEOne of the most important areas of improvement for the next generation of agricultural data and models is to better understand and communicate the inherent uncertainties. This is especially critical in data-limited regions. Here we present a modeling system and its implementation that begins to address these concerns. We demonstrate the ability to simulate broad trends in yields at the county level for sub-annual yields with skills that commensurate previous national/annual level studies.

Regionalization of Root Zone Moisture Estimations from Downscaled Surface Moisture and Environmental Data with the Soil Moisture Analytical Relationship Model

Root-zone soil moisture (RZSM) plays a key role in the hydrologic cycle and regulates water–heat exchange. Although site observations can provide soil profile moisture measurements, they have a restricted representation. Satellites can determine soil moisture on a large scale, yet the depth of detection is limited. RZSM can be estimated on a large scale using the soil moisture analytical relationship (SMAR) and surface soil moisture (SSM). However, the applicability of the SMAR to different deep-root zones and covariate sources is unclear. This paper investigates the applicability of the SMAR in the Shandian River Basin, upstream of the Luan River in China, by combining site and regional soil moisture, soil properties, and meteorological data. In particular, we first compared the estimation results of the SMAR at different depths (10–20 cm; 10–50 cm) and using covariates from different sources (dataset, SMAR-P1; literature, SMAR-P2) at the site in order to generate SMAR calibration parameters. The parameters were then regionalized based on multiple linear regression by combining the SMAR-P1, SMAR-P2, and SMAR calibration parameters in the 10–50 cm root zone. Finally, the Shandian River RZSM was estimated using regional surface soil moisture and the aforementioned regionalized parameters. At the site scale, diffusion coefficient b obtained in the 10–20 cm root zone at the same depth as the surface layer exceeded the upper limit of the SMAR by one. This is not fit an environment within the site context, and thus the SMAR is not applicable at this particular depth. The opposite is observed for the 10–50 cm root zone. In addition, SMAR-P1 (RMSE = 0.02) outperformed SMAR-P2 (RMSE = 0.04) in the estimation of the RZSM at 10–50 cm. Parameter regionalization analysis revealed the failure of SMAR-P2 to pass the significance test (p &gt; 0.05) for building a multivariate linear model, while SMAR-P1 successfully passed the significance test (p &lt; 0.05) and finished the parameter regionalization process. The median RMSE and median R2adj of the regional RZSM results were determined as 0.12 and 0.3, respectively. The regional RZSM agrees with the spatial trend of the Shandian River. This study examines the suitability of the SMAR model in varying deep-root zones and with diverse covariate sources. The results provide a crucial basis for future utilization of the SMAR.

Ensemble of optimised machine learning algorithms for predicting surface soil moisture content at a global scale

Abstract. Accurate information on surface soil moisture (SSM) content at a global scale under different climatic conditions is important for hydrological and climatological applications. Machine-learning-based systematic integration of in situ hydrological measurements, complex environmental and climate data, and satellite observation facilitate the generation of reliable data products to monitor and analyse the exchange of water, energy, and carbon in the Earth system at a proper space–time resolution. This study investigates the estimation of daily SSM using 8 optimised machine learning (ML) algorithms and 10 ensemble models (constructed via model bootstrap aggregating techniques and five-fold cross-validation). The algorithmic implementations were trained and tested using International Soil Moisture Network (ISMN) data collected from 1722 stations distributed across the world. The result showed that the K-neighbours Regressor (KNR) had the lowest root-mean-square error (0.0379 cm3 cm−3) on the “test_random” set (for testing the performance of randomly split data during training), the Random Forest Regressor (RFR) had the lowest RMSE (0.0599 cm3 cm−3) on the “test_temporal” set (for testing the performance on the period that was not used in training), and AdaBoost (AB) had the lowest RMSE (0.0786 cm3 cm−3) on the “test_independent-stations” set (for testing the performance on the stations that were not used in training). Independent evaluation on novel stations across different climate zones was conducted. For the optimised ML algorithms, the median RMSE values were below 0.1 cm3 cm−3. GradientBoosting (GB), Multi-layer Perceptron Regressor (MLPR), Stochastic Gradient Descent Regressor (SGDR), and RFR achieved a median r score of 0.6 in 12, 11, 9, and 9 climate zones, respectively, out of 15 climate zones. The performance of ensemble models improved significantly, with the median RMSE value below 0.075 cm3 cm−3 for all climate zones. All voting regressors achieved r scores of above 0.6 in 13 climate zones; BSh (hot semi-arid climate) and BWh (hot desert climate) were the exceptions because of the sparse distribution of training stations. The metric evaluation showed that ensemble models can improve the performance of single ML algorithms and achieve more stable results. Based on the results computed for three different test sets, the ensemble model with KNR, RFR and Extreme Gradient Boosting (XB) performed the best. Overall, our investigation shows that ensemble machine learning algorithms have a greater capability with respect to predicting SSM compared with the optimised or base ML algorithms; this indicates their huge potential applicability in estimating water cycle budgets, managing irrigation, and predicting crop yields.

Estimation of the biomechanical mammographic deformation of the breast using machine learning models

BackgroundA typical problem in the registration of MRI and X-ray mammography is the nonlinear deformation applied to the breast during mammography. We have developed a method for virtual deformation of the breast using a biomechanical model automatically constructed from MRI. The virtual deformation is applied in two steps: unloaded state estimation and compression simulation. The finite element method is used to solve the deformation process. However, the extensive computational cost prevents its usage in clinical routine. MethodsWe propose three machine learning models to overcome this problem: an extremely randomized tree (first model), extreme gradient boosting (second model), and deep learning-based bidirectional long short-term memory with an attention layer (third model) to predict the deformation of a biomechanical model. We evaluated our methods with 516 breasts with realistic compression ratios up to 76%. FindingsWe first applied one-fold validation, in which the second and third models performed better than the first model. We then applied ten-fold validation. For the unloaded state estimation, the median RMSE for the second and third models is 0.8 mm and 1.2 mm, respectively. For the compression, the median RMSE is 3.4 mm for both models. We evaluated correlations between model accuracy and characteristics of the clinical datasets such as compression ratio, breast volume, and tissue types. InterpretationUsing the proposed models, we achieved accurate results comparable to the finite element model, with a speedup of factor 240 using the extreme gradient boosting model. These proposed models can replace the finite element model simulation, enabling clinically relevant real-time application.

Predicting daily solar radiation using a novel hybrid long short-term memory network across four climate regions of China

Accurate daily solar radiation (Rs) data is crucial for the growth and development of crops, but its direct observation is relatively scarce in most regions around the world. Therefore, accurately estimating daily Rs is of great significance. Regarding the above practical requirement, this study developed a novel hybrid deep learning model (MEA-DBN-LSTM), which combined the Long Short-Term Memory model (LSTM), Deep Belief Network (DBN), and Mind Evolutionary Algorithm (MEA), to estimate the Rs across four different climate regions of China using five easily accessible meteorological input combination (sunshine-based, temperature-based and three complex-based). The results showed that MEA-DBN-LSTM was proved to be the most recommended Rs estimated model across four different climate regions of China, and the corresponding median R2, NSE, RMSE, MAE, and MAPE values ranged 0.805–0.999, 0.656–0.954, 1.069–4.289 MJ m−2 d–1, 0.889–3.532 MJ m−2 d–1, and 0.055–0.296 respectively. Correspondingly, compared to the empirical Rs model, the improvement values of MEA-DBN-LSTM were 12.33–52.13 % for R2, 14.80–69.38 % for NSE, 33.75–69.77 % for RMSE, 36.03–62.70 % for MAE, and 42.40–66.67 % for MAPE, respectively. Moreover, the MEA-DBN-LSTM exhibits improvements in R2, NSE, RMSE, MAE, and MAPE, ranging 1.40–13.50 %, 1.27–24.79 %, 1.81–61.18 %, 1.90–61.30 %, and 2.31–57.55 %, respectively, compared to other forms of LSTM-type models. Specifically, this study found the DBN modules significantly improved the complex-based single LSTM model, but without impact on sunshine-based and temperature-based single LSTM models. Regarding the MEA modules, it was effective in the hyperparameter optimization compared to DBN-LSTM. Furthermore, significant accuracy discrepancies exist among Rs models using different input combinations across four different climate regions, and the complex-based was generally superior to other input combinations. Overall, our findings can provide accurate daily Rs across four different climate regions of China using easily accessible meteorological data, which is of great significance for crop water consumption estimation, agricultural water resources management, crop growth conditions optimization, and sustainable agricultural development.

A comprehensive assessment of in situ and remote sensing soil moisture data assimilation in the APSIM model for improving agricultural forecasting across the US Midwest

Abstract. Today, the most popular approaches in agricultural forecasting leverage process-based crop models, crop monitoring data, and/or remote sensing imagery. Individually, each of these tools has its own unique advantages but is, nonetheless, limited in prediction accuracy, precision, or both. In this study we integrate in situ and remote sensing (RS) soil moisture observations with APSIM model through sequential data assimilation to evaluate the improvement in model predictions of downstream state variables across five experimental sites in the US Midwest. Four RS data products and in situ observations spanning 19 site years were used through two data assimilation approaches, namely ensemble Kalman filter (EnKF) and generalized ensemble filter (GEF), to constrain model states at observed time steps and estimate joint background and observation error matrices. Then, the assimilation's impact on estimates of soil moisture, yield, normalized difference vegetation index (NDVI), tile drainage, and nitrate leaching was assessed across all site years. When assimilating in situ observations, the accuracy of soil moisture forecasts in the assimilation layers was improved by reducing RMSE by an average of 17 % for 10 cm and ∼28 % for 20 cm depth soil layer across all site years. These changes also led to improved simulation of soil moisture in deeper soil layers by an average of 12 %. Although crop yield was improved by an average of 23 %, the greatest improvement in yield accuracy was demonstrated in site years with higher water stress, where assimilation served to increase available soil water for crop uptake. Alternatively, estimates of annual tile drainage and nitrate leaching were not well constrained across the study sites. Trends in drainage constraint suggest the importance of including additional data constraint such as evapotranspiration. The assimilation of RS soil moisture showed a weaker constraint of downstream model state variables when compared to the assimilation of in situ soil moisture. The median reduction in soil moisture RMSE for observed soil layers was lower, on average, by a factor of 5. However, crop yield estimates were still improved overall with a median RMSE reduction of 17.2 %. Crop yield prediction was improved when assimilating both in situ and remote sensing soil moisture observations, and there is strong evidence that yield improvement was higher when under water-stressed conditions. Comparisons of system performance across different combinations of remote sensing data products indicated the importance of high temporal resolution and accurate observation uncertainty estimates when assimilating surface soil moisture observations.

PH-dependent solubility prediction for optimized drug absorption and compound uptake by plants.

Aqueous solubility is the most important physicochemical property for agrochemical and drug candidates and a prerequisite for uptake, distribution, transport, and finally the bioavailability in living species. We here present the first-ever direct machine learning models for pH-dependent solubility in water. For this, we combined almost 300000 data points from 11 solubility assays performed over 24years and over one million data points from lipophilicity and melting point experiments. Data were split into three pH-classes - acidic, neutral and basic - , representing the conditions of stomach and intestinal tract for animals and humans, and phloem and xylem for plants. We find that multi-task neural networks using ECFP-6 fingerprints outperform baseline random forests and single-task neural networks on the individual tasks. Our final model with three solubility tasks using the pH-class combined data from different assays and five helper tasks results in root mean square errors of 0.56 log units overall (acidic 0.61; neutral 0.52; basic 0.54) and Spearman rank correlations of 0.83 (acidic 0.78; neutral 0.86; basic 0.86), making it a valuable tool for profiling of compounds in pharmaceutical and agrochemical research. The model allows for the prediction of compound pH profiles with mean and median RMSE per molecule of 0.62 and 0.56 log units.

DeepAir: Enabling Data-Driven Dynamic Spectrum Sharing via Scalable Forecasting

The rapid uptake of wireless technologies over the past decade has resulted in an increasing pressure on the limited radio spectrum resources. To improve the efficiency of current allocation policies, regulators in many jurisdictions are considering dynamic spectrum sharing. The success, however, of an optimized system hinges on the ability to sense, characterize, and forecast spectrum usage behaviour. Since traditional methods prove unable to scale to a wide range of channels, we propose DeepAir, a robust and scalable model that is capable of learning and predicting complex temporal and spectral dependencies in multivariate spectrum data. Specifically, we design a Sequence-to-Sequence model that employs an encoder-decoder architecture with two Deep Temporal Convolutional Networks. Using a test set consisting of approximately 900 channels in the Land Mobile Radio bands, we obtain a median RMSE and median MAE of 6.51 and 5.15, respectively. We then apply transfer learning to demonstrate the effectiveness of this model in forecasting patterns from any sensor, regardless of the band, sensitivity, and geographical location. Furthermore, the model exhibits no performance degradation up to three years after training for both short and long forecast horizons. Finally, we use DeepAir to quantify spectrum availability to enhance existing spectrum sharing capabilities.

Impact of Preprocessing on Tree Canopy Cover Modelling: Does Gap-Filling of Landsat Time Series Improve Modelling Accuracy?

Preprocessing of Landsat images is a double-edged sword, transforming the raw data into a useful format but potentially introducing unwanted values with unnecessary steps. Through recovering missing data of satellite images in time series analysis, gap-filling is an important, highly developed, preprocessing procedure, but its necessity and effects in numerous Landsat applications, such as tree canopy cover (TCC) modelling, are rarely examined. We address this barrier by providing a quantitative comparison of TCC modelling using predictor variables derived from Landsat time series that included gap-filling versus those that did not include gap-filling and evaluating the effects that gap-filling has on modelling TCC. With 1-year Landsat time series from a tropical region located in Taita Hills, Kenya, and a reference TCC map in 0–100 scales derived from airborne laser scanning data, we designed comparable random forest modelling experiments to address the following questions: 1) Does gap-filling improve TCC modelling based on time series predictor variables including the seasonal composites (SC), spectral-temporal metrics (STMs), and harmonic regression (HR) coefficients? 2) What is the difference in TCC modelling between using gap-filled pixels and using valid (actual or cloud-free) pixels? Two gap-filling methods, one temporal-based method (Steffen spline interpolation) and one hybrid method (MOPSTM) have been examined. We show that gap-filled predictors derived from the Landsat time series delivered better performance on average than non-gap-filled predictors with the average of median RMSE values for Steffen-filled and MOPSTM-filled SC’s being 17.09 and 16.57 respectively, while for non-gap-filled predictors, it was 17.21. MOPSTM-filled SC is 3.7% better than non-gap-filled SC on RMSE, and Steffen-filled SC is 0.7% better than non-gap-filled SC on RMSE. The positive effects of gap-filling may be reduced when there are sufficient high-quality valid observations to generate a seasonal composite. The single-date experiment suggests that gap-filled data (e.g. RMSE of 16.99, 17.71, 16.24, and 17.85 with 100% gap-filled pixels as training and test datasets for four seasons) may deliver no worse performance than valid data (e.g. RMSE of 15.46, 17.07, 16.31, and 18.14 with 100% valid pixels as training and test datasets for four seasons). Thus, we conclude that gap-filling has a positive effect on the accuracy of TCC modelling, which justifies its inclusion in image preprocessing workflows.

Modeling Patient-Specific Desaturation Patterns in Sleep Apnea.

The physiological mechanisms involved in cardio-respiratory responses to sleep apnea events are not yet fully elucidated. A model-based approach is proposed to analyse the acute desaturation response to obstructive apneas. An integrated model of cardio-respiratory interactions was proposed and parameters were identified, using an evolutionary algorithm, on a database composed of 107 obstructive apneas acquired from 10 patients (HYPNOS clinical study). Unsupervised clustering was applied to the identified parameters in order to characterize the phenotype of each response to obstructive apneas. A close match was observed between simulated oxygen saturation ( SaO2) and experimental SaO2 in all identifications (median RMSE = 1.3892 %). Two clusters of parameters, associated with different dynamics related to sleep apnea and periodic breathing were obtained. The proposed patient and event-specific model-based analysis provides understanding on specific desaturation patterns, consequent to apnea events, with potential applications for personalized diagnosis and treatment.

Octree Representation Improves Data Fidelity of Cardiac CT Images and Convolutional Neural Network Semantic Segmentation of Left Atrial and Ventricular Chambers.

To assess whether octree representation and octree-based convolutional neural networks (CNNs) improve segmentation accuracy of three-dimensional images. Cardiac CT angiographic examinations from 100 patients (mean age, 67 years ± 17 [standard deviation]; 60 men) performed between June 2012 and June 2018 with semantic segmentations of the left ventricular (LV) and left atrial (LA) blood pools at the end-diastolic and end-systolic cardiac phases were retrospectively evaluated. Image quality (root mean square error [RMSE]) and segmentation fidelity (global Dice and border Dice coefficients) metrics of the octree representation were compared with spatial downsampling for a range of memory footprints. Fivefold cross-validation was used to train an octree-based CNN and CNNs with spatial downsampling at four levels of image compression or spatial downsampling. The semantic segmentation performance of octree-based CNN (OctNet) was compared with the performance of U-Nets with spatial downsampling. Octrees provided high image and segmentation fidelity (median RMSE, 1.34 HU; LV Dice coefficient, 0.970; LV border Dice coefficient, 0.843) with a reduced memory footprint (87.5% reduction). Spatial downsampling to the same memory footprint had lower data fidelity (median RMSE, 12.96 HU; LV Dice coefficient, 0.852; LV border Dice coefficient, 0.310). OctNet segmentation improved the border segmentation Dice coefficient (LV, 0.612; LA, 0.636) compared with the highest performance among U-Nets with spatial downsampling (Dice coefficients: LV, 0.579; LA, 0.592). Octree-based representations can reduce the memory footprint and improve segmentation border accuracy.Keywords CT, Cardiac, Segmentation, Supervised Learning, Convolutional Neural Network (CNN), Deep Learning Algorithms, Machine Learning Algorithms© RSNA, 2021.