Diagnostic support in pediatric craniopharyngioma using deep learning.

PURPOSE: We studied a pediatric group of patients with sellar-suprasellar tumors, aiming to develop a convolutional deep learning algorithm for radiological assistance to classify them into their respective cohort. METHODS: T1w and T2w preoperative magnetic resonance images of 226 Chilean patients were collected at the Institute of Neurosurgery Dr. Alfonso Asenjo (INCA), which were divided into three classes: healthy control (68 subjects), craniopharyngioma (58 subjects) and differential sellar/suprasellar tumors (100 subjects). RESULTS: The PPV among classes was 0.828±0.039, and the NPV was 0.919±0.063. Also explainable artificial intelligence (XAI) was used, finding that structures that are relevant during diagnosis and radiological evaluation highly influence the decision-making process of the machine. CONCLUSION: This is the first experience of this kind of study in our institution, and it led to promising results on the task of radiological diagnostic support based on explainable artificial intelligence (AI) and deep learning models.

A robust deep learning detector for sleep spindles and K-complexes: towards population norms.

Sleep spindles (SSs) and K-complexes (KCs) are brain patterns involved in cognitive functions that appear during sleep. Large-scale sleep studies would benefit from precise and robust automatic sleep event detectors, capable of adapting the variability in both electroencephalography (EEG) signals and expert annotation rules. We introduce the Sleep EEG Event Detector (SEED), a deep learning system that outperforms existing approaches in SS and KC detection, reaching an F1-score of 80.5% and 83.7%, respectively, on the MASS2 dataset. SEED transfers well and requires minimal fine-tuning for new datasets and annotation styles. Remarkably, SEED substantially reduces the required amount of annotated data by using a novel pretraining approach that leverages the rule-based detector A7. An analysis of 11,224 subjects revealed that SEED's detections provide better estimates of SS population statistics than existing approaches. SEED is a powerful resource for obtaining sleep-event statistics that could be useful for establishing population norms.

Development of machine learning models to predict gestational diabetes risk in the first half of pregnancy.

BACKGROUND: Early prediction of Gestational Diabetes Mellitus (GDM) risk is of particular importance as it may enable more efficacious interventions and reduce cumulative injury to mother and fetus. The aim of this study is to develop machine learning (ML) models, for the early prediction of GDM using widely available variables, facilitating early intervention, and making possible to apply the prediction models in places where there is no access to more complex examinations. METHODS: The dataset used in this study includes registries from 1,611 pregnancies. Twelve different ML models and their hyperparameters were optimized to achieve early and high prediction performance of GDM. A data augmentation method was used in training to improve prediction results. Three methods were used to select the most relevant variables for GDM prediction. After training, the models ranked with the highest Area under the Receiver Operating Characteristic Curve (AUCROC), were assessed on the validation set. Models with the best results were assessed in the test set as a measure of generalization performance. RESULTS: Our method allows identifying many possible models for various levels of sensitivity and specificity. Four models achieved a high sensitivity of 0.82, a specificity in the range 0.72-0.74, accuracy between 0.73-0.75, and AUCROC of 0.81. These models required between 7 and 12 input variables. Another possible choice could be a model with sensitivity of 0.89 that requires just 5 variables reaching an accuracy of 0.65, a specificity of 0.62, and AUCROC of 0.82. CONCLUSIONS: The principal findings of our study are: Early prediction of GDM within early stages of pregnancy using regular examinations/exams; the development and optimization of twelve different ML models and their hyperparameters to achieve the highest prediction performance; a novel data augmentation method is proposed to allow reaching excellent GDM prediction results with various models.

Going deep into schizophrenia with artificial intelligence.

Despite years of research, the mechanisms governing the onset, relapse, symptomatology, and treatment of schizophrenia (SZ) remain elusive. The lack of appropriate analytic tools to deal with the heterogeneity and complexity of SZ may be one of the reasons behind this situation. Deep learning, a subfield of artificial intelligence (AI) inspired by the nervous system, has recently provided an accessible way of modeling and analyzing complex, high-dimensional, nonlinear systems. The unprecedented accuracy of deep learning algorithms in classification and prediction tasks has revolutionized a wide range of scientific fields and is rapidly permeating SZ research. Deep learning has the potential of becoming a valuable aid for clinicians in the prediction, diagnosis, and treatment of SZ, especially in combination with principles from Bayesian statistics. Furthermore, deep learning could become a powerful tool for uncovering the mechanisms underlying SZ thanks to a growing number of techniques designed for improving model interpretability and causal reasoning. The purpose of this article is to introduce SZ researchers to the field of deep learning and review its latest applications in SZ research. In general, existing studies have yielded impressive results in classification and outcome prediction tasks. However, methodological concerns related to the assessment of model performance in several studies, the widespread use of small training datasets, and the little clinical value of some models suggest that some of these results should be taken with caution.

Sleep-spindle identification on EEG signals from polysomnographie recordings using correntropy.

Sleep spindles (SSs) are characteristic electroencephalographic (EEG) waveforms of sleep stages N2 and N3. One of the main problems associated with SS detection is the high number of false positives. In this paper we propose a new periodogram based on correntropy to detect SSs and enhance their characterization. Correntropy is a generalized correlation, under the information theoretic learning framework. A non-negative matrix factorization decomposition of correntropy allows us to obtain a new periodogram, which shows an improved resolution capability compared to the conventional power spectrum density. Preliminary results show that the proposed method obtained a sensitivity rate of 0.868 with a false positive rate of 0.121.

Automated detection of rapid eye movements in children.

We present an automated multiple-step tool to identify Rapid Eye Movements (REMs) in the polysomnogram, based on modeling expert criteria. It begins by identifying the polysomnogram segments compatible with REMs presence. On these segments, high-energy REMs are identified. Then, vicinity zones around those REMs are defined, and lesser-energy REMs are sought in these vicinities. This strategy has the advantage that it can detect lesser-energy REMs without increasing much the false positive detections. Signal processing, feature extraction, and fuzzy logic tools are used to achieve the goal. The tool was trained and validated on a database consisting of 20 all-night polysomnogram recordings (160 hr) of healthy ten-year-old children. Preliminary results on the validation set show 85.5% sensitivity and a false positive rate of 16.2%. Our tool works on complete polysomnogram recordings, without the need of preprocessing, prior knowledge of the hypnogram, or noise-free segments selection.

Registration of renal SPECT and 2.5D US images.

Image registration is the process of transforming different image data sets of an object into the same coordinate system. This is a relevant task in the field of medical imaging; one of its objectives is to combine information from different imaging modalities. The main goal of this study is the registration of renal SPECT (Single Photon Emission Computerized Tomography) images and a sparse set of ultrasound slices (2.5D US), combining functional and anatomical information. Registration is performed after kidney segmentation in both image types. The SPECT segmentation is achieved using a deformable model based on a simplex mesh. The 2.5D US image segmentation is carried out in each of the 2D slices employing a deformable contour and Gabor filters to capture multi-scale image features. Moreover, a renal medulla detection method was developed to improve the US segmentation. A nonlinear optimization algorithm is used for the registration. In this process, movements caused by patient breathing during US image acquisition are also corrected. Only a few reports describe registration between SPECT images and a sparse set of US slices of the kidney, and they usually employ an optical localizer, unlike our method, that performs movement correction using information only from the SPECT and US images. Moreover, it does not require simultaneous acquisition of both image types. The registration method and both segmentations were evaluated separately. The SPECT segmentation was evaluated qualitatively by medical experts, obtaining a score of 5 over a scale from 1 to 5, where 5 represents a perfect segmentation. The 2.5D US segmentation was evaluated quantitatively, by comparing our method with an expert manual segmentation, and obtaining an average error of 3.3mm. The registration was evaluated quantitatively and qualitatively. Quantitatively the distance between the manual segmentation of the US images and the model extracted from the SPECT image was measured, obtaining an average distance of 1.07 pixels on 7 exams. The qualitative evaluation was carried out by a group of physicians who assessed the perceived clinical usefulness of the image registration, rating each registration on a scale from 1 to 5. The average score obtained was 4.1, i.e. relevantly useful for medical purposes.

Automated sleep-spindle detection in healthy children polysomnograms.

We present a new methodology to detect and characterize sleep spindles (SSs), based on the nonlinear algorithms, empirical-mode decomposition, and Hilbert-Huang transform, which provide adequate temporal and frequency resolutions in the electroencephalographic analysis. In addition, the application of fuzzy logic allows to emulate expert's procedures. Additionally, we built a database of 56 all-night polysomnographic recordings from children for training and testing, which is among the largest annotated databases published on the subject. The database was split into training (27 recordings), validation (10 recordings), and testing (19 recordings) datasets. The SS events were marked by sleep experts using visual inspection, and these marks were used as golden standard. The overall SS detection performance on the testing dataset of continuous all-night sleep recordings was 88.2% sensitivity, 89.7% specificity, and 11.9% false-positive (FP) rate. Considering only non-REM sleep stage 2, the results showed 92.2% sensitivity, 90.1% specificity, and 8.9% FP rate. In general, our system presents enhanced results when compared with most systems found in the literature, thus improving SS detection precision significantly without the need of hypnogram information.

Normalized mutual information feature selection.

A filter method of feature selection based on mutual information, called normalized mutual information feature selection (NMIFS), is presented. NMIFS is an enhancement over Battiti's MIFS, MIFS-U, and mRMR methods. The average normalized mutual information is proposed as a measure of redundancy among features. NMIFS outperformed MIFS, MIFS-U, and mRMR on several artificial and benchmark data sets without requiring a user-defined parameter. In addition, NMIFS is combined with a genetic algorithm to form a hybrid filter/wrapper method called GAMIFS. This includes an initialization procedure and a mutation operator based on NMIFS to speed up the convergence of the genetic algorithm. GAMIFS overcomes the limitations of incremental search algorithms that are unable to find dependencies between groups of features.

Extracting fuzzy rules from polysomnographic recordings for infant sleep classification.

A neuro-fuzzy classifier (NFC) of sleep-wake states and stages has been developed for healthy infants of ages 6 mo and onward. The NFC takes five input patterns previously identified on 20-s epochs from polysomnographic recordings and assigns them to one out of five possible classes: Wakefulness, REM-Sleep, Non-REM Sleep Stage 1, Stage 2, and Stage 3-4. The definite criterion for a sleep state or stage to be established is duration of at least 1 min. The data set consisted of a total of 14 continuous recordings of naturally occurring naps (average duration: 143 +/- 39 min), corresponding to a total of 6021 epochs. They were divided in a training, a validation and a test set with 7, 2, and 5 recordings, respectively. During supervised training, the system determined the fuzzy concepts associated to the inputs and the rules required for performing the classification, extracting knowledge from the training set, and pruning nonrelevant rules. Results on an independent test set achieved 83.9 +/- 0.4% of expert agreement. The fuzzy rules obtained from the training examples without a priori information showed a high level of coincidence with the crisp rules stated by the experts, which are based on internationally accepted criteria. These results show that the NFC can be a valuable tool for implementing an automated sleep-wake classification system.

Online data visualization using the neural gas network.

A high-quality distance preserving output representation is provided to the neural gas (NG) network. The nonlinear mapping is determined concurrently along with the codebook vectors. The adaptation rule for codebook positions in the projection space minimizes a cost function that favors the trustworthy preservation of the local topology. The proposed visualization method, called OVI-NG, is an enhancement over curvilinear component analysis (CCA). The results show that the mapping quality obtained with OVI-NG outperforms the original CCA, in terms of the trustworthiness, continuity, topographic function and topology preservation measures.

Cross-entropy embedding of high-dimensional data using the neural gas model.

A cross-entropy approach to mapping high-dimensional data into a low-dimensional space embedding is presented. The method allows to project simultaneously the input data and the codebook vectors, obtained with the Neural Gas (NG) quantizer algorithm, into a low-dimensional output space. The aim of this approach is to preserve the relationship defined by the NG neighborhood function for each pair of input and codebook vectors. A cost function based on the cross-entropy between input and output probabilities is minimized by using a Newton-Raphson method. The new approach is compared with Sammon's non-linear mapping (NLM) and the hierarchical approach of combining a vector quantizer such as the self-organizing feature map (SOM) or NG with the NLM recall algorithm. In comparison with these techniques, our method delivers a clear visualization of both data points and codebooks, and it achieves a better mapping quality in terms of the topology preservation measure q(m).