Madalina Fiterau

Deep Neural Decision Forests

We present Deep Neural Decision Forests - a novel approach that unifies classification trees with the representation learning functionality known from deep convolutional networks, by training them in an end-to-end manner. To combine these two worlds, we introduce a stochastic and differentiable decision tree model, which steers the representation learning usually conducted in the initial layers of a (deep) convolutional network. Our model differs from conventional deep networks because a decision forest provides the final predictions and it differs from conventional decision forests since we propose a principled, joint and global optimization of split and leaf node parameters. We show experimental results on benchmark machine learning datasets like MNIST and ImageNet and find on-par or superior results when compared to state-of-the-art deep models. Most remarkably, we obtain Top5-Errors of only 7.84%/6.38% on ImageNet validation data when integrating our forests in a single-crop, single/seven model GoogLeNet architecture, respectively. Thus, even without any form of training data set augmentation we are improving on the 6.67% error obtained by the best GoogLeNet architecture (7 models, 144 crops).

Using Supervised Machine Learning to Classify Real Alerts and Artifact in Online Multisignal Vital Sign Monitoring Data.

Objective: The use of machine-learning algorithms to classify alerts as real or artifacts in online noninvasive vital sign data streams to reduce alarm fatigue and missed true instability. Design: Observational cohort study. Setting: Twenty-four–bed trauma step-down unit. Patients: Two thousand one hundred fifty-three patients. Intervention: Noninvasive vital sign monitoring data (heart rate, respiratory rate, peripheral oximetry) recorded on all admissions at 1/20 Hz, and noninvasive blood pressure less frequently, and partitioned data into training/validation (294 admissions; 22,980 monitoring hours) and test sets (2,057 admissions; 156,177 monitoring hours). Alerts were vital sign deviations beyond stability thresholds. A four-member expert committee annotated a subset of alerts (576 in training/validation set, 397 in test set) as real or artifact selected by active learning, upon which we trained machine-learning algorithms. The best model was evaluated on test set alerts to enact online alert classification over time. Measurements and Main Results: The Random Forest model discriminated between real and artifact as the alerts evolved online in the test set with area under the curve performance of 0.79 (95% CI, 0.67–0.93) for peripheral oximetry at the instant the vital sign first crossed threshold and increased to 0.87 (95% CI, 0.71–0.95) at 3 minutes into the alerting period. Blood pressure area under the curve started at 0.77 (95% CI, 0.64–0.95) and increased to 0.87 (95% CI, 0.71–0.98), whereas respiratory rate area under the curve started at 0.85 (95% CI, 0.77–0.95) and increased to 0.97 (95% CI, 0.94–1.00). Heart rate alerts were too few for model development. Conclusions: Machine-learning models can discern clinically relevant peripheral oximetry, blood pressure, and respiratory rate alerts from artifacts in an online monitoring dataset (area under the curve > 0.87).

Performance of handover for multiple users in heterogeneous wireless networks

Handover solutions ensuring seamless connectivity and high user-perceived quality of service for a given application context are essential for multi-mode wireless devices in heterogeneous wireless network environments. A critical handover step, the network selection decision, is automatically and transparently made in the users terminal, aiming to keep the user “always best connected”. We propose Quantified Adaptive Delay Selection (QADS), a novel multi-user-aware handover algorithm that maintains high quality of service levels for mobile users performing handover in heterogeneous wireless network environments. QADS is a user-centric solution buildt on the IEEE 802.21 Media Independent Handover standard. It addresses the problem of multiple mobile nodes performing network selection independently, using the same selection algorithm. With innovative mechanisms based on adaptive contention and randomization, the algorithm increases overall user-perceived quality of service.

Informative Projection Recovery for Classification, Clustering and Regression

Data driven decision support systems often benefit from human participation to validate outcomes produced by automated procedures. Perceived utility hinges on the systems ability to learn transparent, comprehensible models from data. We introduce and formalize Informative Projection Recovery: the problem of extracting a set of low-dimensional projections of data which jointly form an accurate solution to a given learning task. We approach this problem with RIPR: a regression-based algorithm that identifies informative projections by optimizing over a matrix of point-wise loss estimators. It generalizes from our previous algorithm, offering solutions to classification, clustering, and regression tasks. Experiments show that RIPR can discover and leverage structures of informative projections in data, if they exist, while yielding accurate and compact models. It is particularly useful in applications involving multivariate numeric data in which expert assessment of the results is of the essence.

Weakly supervised classification of rare aortic valve malformations using unlabeled cardiac MRI sequences

Recent releases of population-scale biomedical repositories such as the UK Biobank have enabled unprecedented access to prospectively collected medical imaging data. Applying machine learning methods to analyze these data holds great promise in facilitating new insights into the genetic and epidemiological associations between anatomical structures and human health. However, the majority of these imaging data are unlabeled and deriving insights is hindered by the cost of manually annotating data at sufficient scale to train state-of-the-art deep learning models. In this work, we develop a weakly supervised deep learning model for Bicuspid Aortic Valve (BAV) classification using up to 4,000 unlabeled cardiac MRI sequences, comprising a total of 120,000 images. Instead of requiring manually labeled training data, weak supervision relies on noisy heuristic functions defined by domain experts to automatically generate large-scale, imperfect training sets. By leveraging new theoretical work on coping with label noise, models can use weaker supervision sources than was previously possible. In our BAV models, this approach substantially outperforms a traditional supervised baseline trained on hand-labeled data alone, with a 64% improvement in mean F1 score (37.8 to 61.4) on held out test data. In a validation experiment using 9,230 individuals with MRIs and long-term outcome data from the UK Biobank, applying the best-performing BAV classification model identified a subset of individuals with a 1.8-fold increase in risk of a major adverse cardiac event (p <0.001). This work formalizes the first deep learning baseline for aortic valve classification and outlines a general strategy for using weak supervision to analyze large collections of unlabeled medical images. Author summary We developed a deep learning model for Bicuspid Aortic Valve (BAV) classification using up to 4,000 unlabeled cardiac MRI sequences, comprising a total of 120,000 images. Instead of requiring manually labeled training data, as is typical in machine learning, our approach relies on noisy heuristic functions defined by domain experts to automatically generate large-scale, imperfect training sets. In our experiments, this approach substantially outperforms a traditional supervised baseline trained on hand-labeled data alone. In a validation experiment using 9,230 individuals with MRIs and long-term outcome data from the UK Biobank, applying the best-performing BAV classification model identified a subset of individuals with a 1.8-fold increase in risk of a major adverse cardiac event. This work formalizes the first deep learning baseline for aortic valve classification and outlines a general strategy for using weak supervision to analyze large collections of unlabeled medical images.

Machine learning in human movement biomechanics: Best practices, common pitfalls, and new opportunities

Traditional laboratory experiments, rehabilitation clinics, and wearable sensors offer biomechanists a wealth of data on healthy and pathological movement. To harness the power of these data and make research more efficient, modern machine learning techniques are starting to complement traditional statistical tools. This survey summarizes the current usage of machine learning methods in human movement biomechanics and highlights best practices that will enable critical evaluation of the literature. We carried out a PubMed/Medline database search for original research articles that used machine learning to study movement biomechanics in patients with musculoskeletal and neuromuscular diseases. Most studies that met our inclusion criteria focused on classifying pathological movement, predicting risk of developing a disease, estimating the effect of an intervention, or automatically recognizing activities to facilitate out-of-clinic patient monitoring. We found that research studies build and evaluate models inconsistently, which motivated our discussion of best practices. We provide recommendations for training and evaluating machine learning models and discuss the potential of several underutilized approaches, such as deep learning, to generate new knowledge about human movement. We believe that cross-training biomechanists in data science and a cultural shift toward sharing of data and tools are essential to maximize the impact of biomechanics research.

Semi automated adjudication of vital sign alerts in step-down units

Machine Learning (ML) has shown predictive utility in analyzing vital sign (VS) data collected from physiologically unstable monitored patients. Training an ML model usually requires sizable amounts of labeled ground-truth data typically obtained via laborious manual chart reviews by expert clinicians.