Independent Validation Data Research Articles

Abstract 5359: Translating state-of-the-art deep learning predictions of treatment efficacy to clinical practice

Abstract Despite the clinical success of immune checkpoint blockade therapies, in many patients the disease becomes unresponsive or resistant to treatment. Attempts to predict treatment efficacy and identify determinants of response have suffered from limited accuracy. Here, we introduce a state-of-the-art deep learning method that better captures the complex relationships among the host, the immune system, and tumor biology. Our method, an explainable deep learning framework, is based on transformer architecture that combines data with different feature sets and clinical endpoints for survival prediction or classification. The framework includes(1) a new loss function based on Harrell’s concordance-index, (2) an explainability module providing importance score between features based on mutual contribution to predictions, and (3) a transfer learning strategy that enables leveraging diverse clinical data sets in the public or private domain. Using seven data sets comprising more than 140,000 patients from immunooncologic (IO), targeted, and chemotherapy treatments, our method consistently outperformed other methods previously described in the literature, including CoxPH and random survival forest (RF). For example, our framework achieved a concordance index of 0.66 (±0.04) vs. 0.60 (±0.04) of the second-best method (RF) on MYSTIC IO arms using only clinical data. Utilizing our explainability module, we identified key features driving response prediction that are consistent with those in previous publications. For example, in the Chowell et al. data set, we identified albumin, neutrophil-to-lymphocyte ratio, prior-chemo-treatment, and tumor mutation burden as the most important features. We also identified in sparse mutation calls of 469 genes from the Samstein et al. data set, functional modules of several genes, each with a strong predictive power. For example, the functional module with the highest hazard ratio (HR) (1.43) was the gene pair KEAP1 and STK11, which are well-established drivers of resistance to multiple therapies. We further validated these results across multiple data sets: The Cancer Genome Atlas HR = 1.23), GENIE/anti-PD-L1 (HR = 1.06), and MYSTIC/anti-PD-L1/CTLA-4 (HR = 1.39). Another example is a functional module related to adaptive immunity, comprising the genes AKT2, BTK, CDC73, HLA-B, IKBKE, INPPL1, RFWD2, TRAF2, and WHSC1. This module stratified patients with an HR of 0.58 within the Samstein et al. data set and 0.42 within the independent validation data set. We describe a new deep learning method with state-of-the-art performance in survival prediction and the potential to uncover biological and clinical insights related to disease response and resistance. Our framework simplifies the process of translating complex AI models to clinical practice and may accelerate the identification of targetable drivers of resistance and response. Citation Format: Gustavo Arango-Argoty, Etai Jacob. Translating state-of-the-art deep learning predictions of treatment efficacy to clinical practice. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5359.

Clusters of co-abundant proteins in the brain cortex associated with fronto-temporal lobar degeneration

BackgroundFrontotemporal lobar degeneration (FTLD) is characterized pathologically by neuronal and glial inclusions of hyperphosphorylated tau or by neuronal cytoplasmic inclusions of TDP43. This study aimed at deciphering the molecular mechanisms leading to these distinct pathological subtypes.MethodsTo this end, we performed an unbiased mass spectrometry-based proteomic and systems-level analysis of the middle frontal gyrus cortices of FTLD-tau (n = 6), FTLD-TDP (n = 15), and control patients (n = 5). We validated these results in an independent patient cohort (total n = 24).ResultsThe middle frontal gyrus cortex proteome was most significantly altered in FTLD-tau compared to controls (294 differentially expressed proteins at FDR = 0.05). The proteomic modifications in FTLD-TDP were more heterogeneous (49 differentially expressed proteins at FDR = 0.1). Weighted co-expression network analysis revealed 17 modules of co-regulated proteins, 13 of which were dysregulated in FTLD-tau. These modules included proteins associated with oxidative phosphorylation, scavenger mechanisms, chromatin regulation, and clathrin-mediated transport in both the frontal and temporal cortex of FTLD-tau. The most strongly dysregulated subnetworks identified cyclin-dependent kinase 5 (CDK5) and polypyrimidine tract-binding protein 1 (PTBP1) as key players in the disease process. Dysregulation of 9 of these modules was confirmed in independent validation data sets of FLTD-tau and control temporal and frontal cortex (total n = 24). Dysregulated modules were primarily associated with changes in astrocyte and endothelial cell protein abundance levels, indicating pathological changes in FTD are not limited to neurons.ConclusionsUsing this innovative workflow and zooming in on the most strongly dysregulated proteins of the identified modules, we were able to identify disease-associated mechanisms in FTLD-tau with high potential as biomarkers and/or therapeutic targets.

Multianalyte Serum Biomarker Panel for Early Detection of Pancreatic Adenocarcinoma.

We determined whether a large, multianalyte panel of circulating biomarkers can improve detection of early-stage pancreatic ductal adenocarcinoma (PDAC). We defined a biologically relevant subspace of blood analytes on the basis of previous identification in premalignant lesions or early-stage PDAC and evaluated each in pilot studies. The 31 analytes that met minimum diagnostic accuracy were measured in serum of 837 subjects (461 healthy, 194 benign pancreatic disease, and 182 early-stage PDAC). We used machine learning to develop classification algorithms using the relationship between subjects on the basis of their changes across the predictors. Model performance was subsequently evaluated in an independent validation data set from 186 additional subjects. A classification model was trained on 669 subjects (358 healthy, 159 benign, and 152 early-stage PDAC). Model evaluation on a hold-out test set of 168 subjects (103 healthy, 35 benign, and 30 early-stage PDAC) yielded an area under the receiver operating characteristic curve (AUC) of 0.920 for classification of PDAC from non-PDAC (benign and healthy controls) and an AUC of 0.944 for PDAC versus healthy controls. The algorithm was then validated in 146 subsequent cases presenting with pancreatic disease (73 benign pancreatic disease and 73 early- and late-stage PDAC cases) and 40 healthy control subjects. The validation set yielded an AUC of 0.919 for classification of PDAC from non-PDAC and an AUC of 0.925 for PDAC versus healthy controls. Individually weak serum biomarkers can be combined into a strong classification algorithm to develop a blood test to identify patients who may benefit from further testing.

A plasma 3-marker microRNA biosignature distinguishes spinal tuberculosis from other spinal destructive diseases and pulmonary tuberculosis.

Accurate spinal tuberculosis (TB) diagnosis is of utmost importance for adequately treating and managing the disease. Given the need for additional diagnostic tools, this study aimed to investigate the utility of host serum miRNA biomarkers for diagnosing and distinguishing spinal tuberculosis (STB) from pulmonary tuberculosis (PTB) and other spinal diseases of different origins (SDD). For a case-controlled investigation, a total of 423 subjects were voluntarily recruited, with 157 cases of STB, 83 cases of SDD, 30 cases of active PTB, and 153 cases of healthy controls (CONT) in 4 clinical centers. To discover the STB-specific miRNA biosignature, a high-throughput miRNA profiling study was performed in the pilot study with 12 cases of STB and 8 cases of CONT using the Exiqon miRNA PCR array platform. A bioinformatics study identified that the 3-plasma miRNA combination (hsa-miR-506-3p, hsa-miR-543, hsa-miR-195-5p) might serve as a candidate biomarker for STB. The subsequent training study developed the diagnostic model using multivariate logistic regression in training data sets, including CONT(n=100) and STB (n=100). Youden's J index determined the optimal classification threshold. Receiver Operating Characteristic (ROC) curve analysis showed that 3-plasma miRNA biomarker signatures have an area under the curve (AUC) = 0.87, sensitivity = 80.5%, and specificity = 80.0%. To explore the possible potential to distinguish spinal TB from PDB and other SDD, the diagnostic model with the same classification threshold was applied to the analysis of the independent validation data set, including CONT(n=45), STB(n=45), brucellosis spondylitis (BS, n=30), PTB (n=30), spinal tumor (ST, n=30) and pyogenic spondylitis (PS, n=23). The results showed diagnostic model based on three miRNA signatures could discriminate the STB from other SDD groups with sensitivity=80%, specificity=96%, Positive Predictive Value (PPV)=84%, Negative Predictive Value (NPV)=94%, the total accuracy rate of 92%. These results indicate that this 3-plasma miRNA biomarker signature could effectively discriminate the STB from other spinal destructive diseases and pulmonary tuberculosis. The present study shows that the diagnostic model based on 3-plasma miRNA biomarker signature (hsa-miR-506-3p, hsa-miR-543, hsa-miR-195-5p) may be used for medical guidance to discriminate the STB from other spinal destructive disease and pulmonary tuberculosis.

New model of glucose-insulin regulation characterizes effects of physical activity and facilitates personalized treatment evaluation in children and adults with type 1 diabetes.

Accurate treatment adjustment to physical activity (PA) remains a challenging problem in type 1 diabetes (T1D) management. Exercise-driven effects on glucose metabolism depend strongly on duration and intensity of the activity, and are highly variable between patients. In-silico evaluation can support the development of improved treatment strategies, and can facilitate personalized treatment optimization. This requires models of the glucose-insulin system that capture relevant exercise-related processes. We developed a model of glucose-insulin regulation that describes changes in glucose metabolism for aerobic moderate- to high-intensity PA of short and prolonged duration. In particular, we incorporated the insulin-independent increase in glucose uptake and production, including glycogen depletion, and the prolonged rise in insulin sensitivity. The model further includes meal absorption and insulin kinetics, allowing simulation of everyday scenarios. The model accurately predicts glucose dynamics for varying PA scenarios in a range of independent validation data sets, and full-day simulations with PA of different timing, duration and intensity agree with clinical observations. We personalized the model on data from a multi-day free-living study of children with T1D by adjusting a small number of model parameters to each child. To assess the use of the personalized models for individual treatment evaluation, we compared subject-specific treatment options for PA management in replay simulations of the recorded data with altered meal, insulin and PA inputs.

Machine learning based identification of structural brain alterations underlying suicide risk in adolescents

Suicide is the third leading cause of death for individuals between 15 and 19 years of age. The high suicide mortality rate and limited prior success in identifying neuroimaging biomarkers indicate that it is crucial to improve the accuracy of clinical neural signatures underlying suicide risk. The current study implements machine-learning (ML) algorithms to examine structural brain alterations in adolescents that can discriminate individuals with suicide risk from typically developing (TD) adolescents at the individual level. Structural MRI data were collected from 79 adolescents who demonstrated clinical levels of suicide risk and 79 demographically matched TD adolescents. Region-specific cortical/subcortical volume (CV/SCV) was evaluated following whole-brain parcellation into 1000 cortical and 12 subcortical regions. CV/SCV parameters were used as inputs for feature selection and three ML algorithms (i.e., support vector machine [SVM], K-nearest neighbors, and ensemble) to classify adolescents at suicide risk from TD adolescents. The highest classification accuracy of 74.79% (with sensitivity = 75.90%, specificity = 74.07%, and area under the receiver operating characteristic curve = 87.18%) was obtained for CV/SCV data using the SVM classifier. Identified bilateral regions that contributed to the classification mainly included reduced CV within the frontal and temporal cortices but increased volume within the cuneus/precuneus for adolescents at suicide risk relative to TD adolescents. The current data demonstrate an unbiased region-specific ML framework to effectively assess the structural biomarkers of suicide risk. Future studies with larger sample sizes and the inclusion of clinical controls and independent validation data sets are needed to confirm our findings.

A new score combining compound muscle action potential (CMAP) amplitudes and motor score is predictive of motor outcome after AVXS-101 (Onasemnogene Abeparvovec) SMA therapy

Spinal muscular atrophy 1 (SMA1) is a severe early genetic disease with degeneration of motor neurons. Motor development is still suboptimal after gene replacement therapy in symptomatic patients. In this study, compound muscle action potential (CMAP) amplitudes were explored as predictors of motor recovery after gene therapy. Thirteen symptomatic SMA1 patients were prospectively included at the Necker Enfants Malades Hospital, Paris, France (Cohort 1) and 12 at the other pediatric neuromuscular reference centers of the French Filnemus network (Cohort 2). In Cohort 1, median CMAP amplitudes showed the best improvement between baseline and the 12 months visit compared to the other tested nerves (ulnar, fibular and tibial). High median CMAP amplitudes at baseline was associated with unaided sitting achievement at M6 (AUC 90%). None of the patients with CHOPINTEND at M0 < 30/64 and median CMAP < 0.5 mV achieved unaided sitting at M6 and this result was confirmed on Cohort 2 used as an independent validation data. Thus, median CMAP amplitude is a valid biomarker for routine practice to predict sitting at M6. A median CMAP amplitude over 0.5 mV at baseline may predict better motor recovery.

A new blood based epigenetic age predictor for adolescents and young adults

Children have special rights for protection compared to adults in our society. However, more than 1/4 of children globally have no documentation of their date of birth. Hence, there is a pressing need to develop biological methods for chronological age prediction, robust to differences in genetics, psychosocial events and physical living conditions. At present, DNA methylation is the most promising biological biomarker applied for age assessment. The human genome contains around 28 million DNA methylation sites, many of which change with age. Several epigenetic clocks accurately predict chronological age using methylation levels at age associated GpG-sites. However, variation in DNA methylation increases with age, and there is no epigenetic clock specifically designed for adolescents and young adults. Here we present a novel age Predictor for Adolescents and Young Adults (PAYA), using 267 CpG methylation sites to assess the chronological age of adolescents and young adults. We compared different preprocessing approaches and investigated the effect on prediction performance of the epigenetic clock. We evaluated performance using an independent validation data set consisting of 18-year-old individuals, where we obtained a median absolute deviation of just below 0.7 years. This tool may be helpful in age assessment of adolescents and young adults. However, there is a need to investigate the robustness of the age predictor across geographical and disease populations as well as environmental effects.

Detection of Intracranial Aneurysms Using Multiphase CT Angiography with a Deep Learning Model

Determine the effect of a multiphase fusion deep-learning model with automatic phase selection in detection of intracranial aneurysm (IA) from computed tomography angiography (CTA) images. CTA images of intracranial arteries from patients at Ningbo First Hospital were retrospectively analyzed. Images were randomly classified as training data, internal validation data, or test data. CTA images from cases examined by digital subtraction angiography (DSA) were examined for independent validation. A deep-learning model was constructed by automatic phase selection of multiphase fusion, and compared to the single-phase algorithm to evaluate algorithm sensitivity. We analyzed 1110 patients (1493 aneurysms) as training data, 139 patients (174 aneurysms) as internal validation data, and 134 patients (175 aneurysms) as test data. The sensitivity of the multiphase analysis of the internal validation data, test data, and independent validation data were greater than from the single-phase analysis. The recall of the multiphase selection was greater or equal to that of single-phase selection in the aneurysm position, shape, size, and rupture status. Use of the test data to determine the presence and absence of aneurysm rupture led to a recall from multiphase selection of 94.8% and 87.6% respectively; both of these values were greater than those from single-phase selection (89.6% and 79.4%). A multiphase fusion deep learning model with automatic phase selection provided automated detection of IAs with high sensitivity.

Building a high-resolution site index map using boosted regression trees: the Norwegian case

Accurate estimation of site productivity is essential for forest projections and scenario modelling. We present and evaluate models to predict site index (SI) and whether a site is productive (potential total stem volume production ≥ 1 m3·ha−1·year−1) in a wall-to-wall high-resolution (16 m × 16 m) SI map for Norway. We investigate whether remotely sensed data improve predictions. We also study the advantages and disadvantages of using boosted regression trees (BRT), a machine-learning algorithm, to create high-accuracy SI maps. We use climatic and topographical data, soil parent material, a land resource map, and depth to water, together with Sentinel-2 satellite images and airborne laser scanning metrics, as predictor variables. We use the SI observed at more than 10 000 National Forest Inventory (NFI) sample plots throughout Norway to fit BRT models and validate the models using 5822 independent temporary plots from the NFI. We benchmark our results against SI estimates from forest monitoring inventories. We find that the SI from BRT has root mean squared error (RMSE) ranging from 2.3 m (hardwoods) to 3.6 m (spruce) when tested against independent validation data from the NFI temporary plots. These RMSEs are similar or marginally better than an evaluation of SI estimates from operational forest management plans where SI normally stems from manual photo interpretation.

Evaluation of Model-Based PM2.5 Estimates for Exposure Assessment during Wildfire Smoke Episodes in the Western U.S.

Investigating the health impacts of wildfire smoke requires data on people's exposure to fine particulate matter (PM2.5) across space and time. In recent years, it has become common to use machine learning models to fill gaps in monitoring data. However, it remains unclear how well these models are able to capture spikes in PM2.5 during and across wildfire events. Here, we evaluate the accuracy of two sets of high-coverage and high-resolution machine learning-derived PM2.5 data sets created by Di et al. and Reid et al. In general, the Reid estimates are more accurate than the Di estimates when compared to independent validation data from mobile smoke monitors deployed by the US Forest Service. However, both models tend to severely under-predict PM2.5 on high-pollution days. Our findings complement other recent studies calling for increased air pollution monitoring in the western US and support the inclusion of wildfire-specific monitoring observations and predictor variables in model-based estimates of PM2.5. Lastly, we call for more rigorous error quantification of machine-learning derived exposure data sets, with special attention to extreme events.

Comparing three spatial modeling tools for assessing urban ecosystem services

The (re)integration of nature into cities has been a progressively promoted strategy to foster sustainable urbanization. To quantify the ecosystem services (ES) provided by urban nature, spatially explicit ES modeling is a key method. However, particularly in absence of independent validation data, there is a clear need for multi-model assessments in order to better understand uncertainties in model outcomes. Here we applied three commonly used open-source ES models (i.e., InVEST, LUCI, NC-Model) to quantify three key urban ES (i.e., local temperature regulation, flood protection and global climate regulation) using the city of The Hague (The Netherlands) as a case study. We quantified the three ES for the current situation and under two hypothetical scenarios representing changes in the amount of vegetation within the city. We found mostly positive correlations between the estimates for a given ES (Spearman’s ρ from 0.11 to 0.84). Yet, our comparison also revealed systematic differences in the ES indicator values between the ES models, as well as different responses to the scenarios. These differences may stem from differences in model structure (i.e., differences in biophysical processes accounted for) and model parameterization (i.e., differences in the value used to quantify a given biophysical process). To further advance urban ES modeling, we recommend i) to improve the representation of urban nature (e.g., green roofs, bioswales, gardens) and urban-specific conditions and processes (e.g., drainage systems, building patterns, soil characteristics) in urban ES models and ii) to systematically account for uncertainty in (urban) ES assessments (e.g., through multi-model assessments).

Inter-rater reliability and prognostic value of baseline Radiographic Assessment of Lung Edema (RALE) scores in observational cohort studies of inpatients with COVID-19

ObjectivesTo reliably quantify the radiographic severity of COVID-19 pneumonia with the Radiographic Assessment of Lung Edema (RALE) score on clinical chest X-rays among inpatients and examine the prognostic value of...

Mechanistic model for human brain metabolism and its connection to the neurovascular coupling.

The neurovascular and neurometabolic couplings (NVC and NMC) connect cerebral activity, blood flow, and metabolism. This interconnection is used in for instance functional imaging, which analyses the blood-oxygen-dependent (BOLD) signal. The mechanisms underlying the NVC are complex, which warrants a model-based analysis of data. We have previously developed a mechanistically detailed model for the NVC, and others have proposed detailed models for cerebral metabolism. However, existing metabolic models are still not fully utilizing available magnetic resonance spectroscopy (MRS) data and are not connected to detailed models for NVC. Therefore, we herein present a new model that integrates mechanistic modelling of both MRS and BOLD data. The metabolic model covers central metabolism, using a minimal set of interactions, and can describe time-series data for glucose, lactate, aspartate, and glutamate, measured after visual stimuli. Statistical tests confirm that the model can describe both estimation data and predict independent validation data, not used for model training. The interconnected NVC model can simultaneously describe BOLD data and can be used to predict expected metabolic responses in experiments where metabolism has not been measured. This model is a step towards a useful and mechanistically detailed model for cerebral blood flow and metabolism, with potential applications in both basic research and clinical applications.

The Influence of FY-4A High-Frequency LST Data on Data Assimilation in a Climate Model

Based on the Beijing Climate Center’s land surface model BCC_AVIM2.0, an ensemble Kalman filter (EnKF) algorithm is developed to assimilate the land surface temperature (LST) product of the first satellite of Fengyun-4 series meteorological satellites of China to study the influence of LST data with different time frequencies on the surface temperature data assimilations. The MODIS daytime and nighttime LST products derived from Terra and Aqua satellites are used as independent validation data to test the assimilation results. The results show that diurnal variation information in the FY-4A LST data has significant effect on the assimilation results. When the time frequencies of the assimilated FY-4A LST data are sufficient, the assimilation scheme can effectively reduce the errors and the assimilation results reflect more reasonable spatial and temporal distributions. The assimilation experiments with a 3 h time frequency show less bias as well as RMSEs and higher temporal correlations than that of the model simulations at both daytime and nighttime periods. As the temporal frequency of assimilated LST observations decreases, the assimilation effects gradually deteriorate. When diurnal variation information is not considered at all in the assimilation, the assimilation with 24 h time frequency showed the largest errors and smallest time correlations in all experiments. The results demonstrate the potential of assimilating high-frequency FY-4A LST data to improve the performance of the BCC_AVIM2.0 land surface model. Furthermore, this study indicates that the diurnal variation information is a necessary factor needed to be considered when assimilating the FY-4A LST.

Estimation of lidar-based gridded DEM uncertainty with varying terrain roughness and point density

Light detection and ranging (lidar) scanning systems can be used to provide a point cloud with high quality and point density. Gridded digital elevation models (DEMs) interpolated from laser scanning point clouds are widely used due to their convenience, however, DEM uncertainty is rarely provided. This paper proposes an end-to-end workflow to quantify the uncertainty (i.e., standard deviation) of a gridded lidar-derived DEM. A benefit of the proposed approach is that it does not require independent validation data measured by alternative means. The input point cloud requires per point uncertainty which is derived from lidar system observational uncertainty. The propagated uncertainty caused by interpolation is then derived by the general law of propagation of variances (GLOPOV) with simultaneous consideration of both horizontal and vertical point cloud uncertainties. Finally, the interpolated uncertainty is then scaled by point density and a measure of terrain roughness to arrive at the final gridded DEM uncertainty. The proposed approach is tested with two lidar datasets measured in Waikoloa, Hawaii, and Sitka, Alaska. Triangulated irregular network (TIN) interpolation is chosen as the representative gridding approach. The results indicate estimated terrain roughness/point density scale factors ranging between 1 (in flat areas) and 7.6 (in high roughness areas), with a mean value of 2.3 for the Waikoloa dataset and between 1 and 9.2 with a mean value of 1.2 for the Sitka dataset. As a result, the final gridded DEM uncertainties are estimated between 0.059 m and 0.677 m with a mean value of 0.164 m for the Waikoloa dataset and between 0.059 m and 1.723 m with a mean value of 0.097 m for the Sitka dataset.

Biomarker Discovery for Meta-Classification of Melanoma Metastatic Progression Using Transfer Learning.

Melanoma is considered to be the most serious and aggressive type of skin cancer, and metastasis appears to be the most important factor in its prognosis. Herein, we developed a transfer learning-based biomarker discovery model that could aid in the diagnosis and prognosis of this disease. After applying it to the ensemble machine learning model, results revealed that the genes found were consistent with those found using other methodologies previously applied to the same TCGA (The Cancer Genome Atlas) data set. Further novel biomarkers were also found. Our ensemble model achieved an AUC of 0.9861, an accuracy of 91.05, and an F1 score of 90.60 using an independent validation data set. This study was able to identify potential genes for diagnostic classification (C7 and GRIK5) and diagnostic and prognostic biomarkers (S100A7, S100A7, KRT14, KRT17, KRT6B, KRTDAP, SERPINB4, TSHR, PVRL4, WFDC5, IL20RB) in melanoma. The results show the utility of a transfer learning approach for biomarker discovery in melanoma.

Performance Analysis of the Temperature and Humidity Profiles Retrieval for FY-3D/MWTHS in Arctic Regions

The special geographical location of the polar regions increases the difficulty of modeling surface emissivity, thus the physical retrieval algorithms of the temperature and humidity profiles for microwave radiometers mainly focus on the regions between 60°S and 60°N. In this paper, the deep neural networks (DNN) and long short-term memory (LSTM) models are first implemented to retrieve atmospheric temperature and humidity profiles in real time from FY-3D/MWHTS in Arctic regions and are compared with the physical retrieval algorithm. The hyperparameters of the machine learning models are determined using the grid search and 10-fold cross-validation. Results show that, compared with the physical retrieval algorithm, the retrieval accuracies of the atmospheric temperature and humidity profiles of the DNN and LSTM models in June 2021 are higher over sea ice, and the maximum retrieval accuracies are improved by about 3.5 K and 42%. Over land, the retrieval accuracies of the atmospheric temperature profiles for the DNN and LSTM models in June 2021 are improved by about 5 K. The retrieved humidity results for these two models are not compared with the physical retrieval algorithm, which fails for the humidity profile retrieval over land. In addition, the retrieval results of the DNN-based and LSTM-based models using the independent validation data in February, April, and September are also evaluated over different surface types. The RMSEs of the retrieved temperature profiles for the two models are within 4 K, except for the near-surface, and the humidity profiles are within 25%, except for in February. The temperature profiles in September and the humidity profiles in February are somewhat reduced compared to other months because of the highly variable emissivity properties in autumn and winter. Overall results show that the machine learning method can well-evaluate the retrieval capability of FY-3D/MWHTS of the atmospheric temperature and humidity profiles in Arctic regions.

Comparative Analysis of Photosynthetically Active Radiation Models Based on Radiometric Attributes in Mainland Spain

The aims of this work are to present an analysis of quality solar radiation data and develop several hourly models of photosynthetically active radiation (PAR) using combinations of radiometric variables such as global horizontal irradiance (GHI), diffuse horizontal irradiance (DHI), and direct normal irradiance (DNI) from their dimensionless indices atmospheric clearness index (kt), horizontal diffuse fraction (kd), and normal direct fraction (kb) together with solar elevation angle (α). GHI, DHI, and DNI data with 1-minute frequencies in the period from 2016 to 2021 from CEDER-CIEMAT, in a northern plateau, and PSA-CIEMAT in the southeast of the Iberian Peninsula, were used to compare two locations with very different climates according to the Köppen—Geiger classification. A total of 15 multilinear models were fitted and validated (with independent training and validation data) using first the whole dataset and then by kt intervals. In most cases, models including the clearness index showed better performance, and among them, models that also use the solar elevation angle as a variable obtained remarkable results. Additionally, according to the statistical validation, these models presented good results when they were compared with models in the bibliography. Finally, the model validation statistics indicate a better performance of the interval models than the complete models.

Predicting cognitive decline in Parkinson’s disease using FDG-PET–based supervised learning

BackgroundCognitive impairment is a common symptom of Parkinson’s disease (PD) that increases in risk and severity as the disease progresses. An accurate prediction of the risk of progression from the mild cognitive impairment (MCI) stage to the dementia (PDD) stage is an unmet clinical need.MethodsWe investigated the use of a supervised learning algorithm called the support vector machine (SVM) to retrospectively stratify patients on the basis of brain fluorodeoxyglucose-PET (FDG-PET) scans. Of 43 patients with PD-MCI according to the baseline scan, 23 progressed to PDD within a 5-year period, whereas 20 maintained stable MCI. The baseline scans were used to train a model, which separated patients identified as PDD converters versus those with stable MCI with 95% sensitivity and 91% specificity.ResultsIn an independent validation data set of 19 patients, the AUC was 0.73, with 67% sensitivity and 80% specificity. The SVM model was topographically characterized by hypometabolism in the temporal and parietal lobes and hypermetabolism in the anterior cingulum and putamen and the insular, mesiotemporal, and postcentral gyri. The performance of the SVM model was further tested on 2 additional data sets, which confirmed that the model was also sensitive to later-stage PDD (17 of 19 patients; 89% sensitivity) and dementia with Lewy bodies (DLB) (16 of 17 patients; 94% sensitivity), but not to normal cognition PD (2 of 17 patients). Finally, anti-PD medication status did not change the SVM classification of the other set of 10 patients with PD who were scanned twice, ON and OFF medication.ConclusionsThese results potentially indicate that the proposed FDG-PET–based SVM classifier has utility for providing an accurate prognosis of dementia development in patients with PD-MCI.