Keywords
Artificial intelligence - machine learning - deep learning - feature engineering -
federated learning
Introduction
Enormous volumes of healthcare data combined with improved capabilities to store,
process, and analyze such data have prompted rapid advances in Artificial Intelligence
(AI) techniques in the medical domain. A recent review summarized the growing number
of peer-reviewed publications addressing the development and use of AI algorithms
in medicine and highlighted the significant need for additional advances and evidence
for these innovations to be translated effectively into practice [1].
In 2018, AI researchers tackled some of the most significant challenges in practically
applying AI in health, such as incomplete and inconsistent data and the need to access
and analyze sensitive healthcare data sets across institutions.
Paper Selection Method
The recommended, standardized methodology for selection of best papers for the International
Medical Informatics Association (IMIA) Yearbook [2] was adapted for the special section theme of “AI in Health: New Opportunities, Challenges,
and Practical Implications.” The Ovid MEDLINE® and Web of Science® databases were queried in February 2019 for peer-reviewed, original research articles
that addressed the special section theme and were published in the English language
in 2018. Editorials, letters, and opinion pieces were deliberately excluded from consideration
as AI in health was a popular topic in health journals in 2018, and perspectives were
common and often redundant.
Section editors generated Medical Subject Headings (MeSH®) and keyword search terms that represented state-of-the-art AI techniques and topics,
and queries were refined iteratively for each database. Concepts incorporated included
artificial intelligence, machine learning, deep learning, federated learning, reinforcement
learning, neural network, convolutional network, variational autoencoder, deep segmentation,
Bayesian model, latent variable, feature engineering, network embedding, gene interaction
network, and automated pattern recognition. In PubMed, the exploded MeSH® term “Medical Informatics” was used as a domain filter. In Web of Science®, the search was limited by the Research Area of “Medical Informatics.” Article type
filters were applied to retrieve research articles from journals or published peer-reviewed
proceedings.
The MEDLINE® and Web of Science® queries retrieved 921 and 621 citations, respectively. After removal of duplicates,
1,480 citations remained. Section editors screened this initial list for relevance
to the theme and scientific quality, and they rated each paper as “keep,” “pend,”
or “discard.” Papers rated as “keep” by one of the section editors were independently
reviewed and scored by section editors to yield the top 15 candidate best papers [3]–[17]. Criteria for scoring included innovation beyond established AI techniques, work
that addressed substantial challenges in the field, and rigorous scientific evaluations.
The top-ranked 15 manuscripts underwent formal peer review by at least four internationally
renowned scientists from across the world. Based on ratings and comments from peer
review, section editors’ recommendations, and input from the IMIA Yearbook Editorial
Board, three papers [3], [8], [11] were chosen as 2018 best papers in the AI for Health special section of the IMIA
Yearbook.
Table 1
Best paper selection of articles for the IMIA Yearbook of Medical Informatics 2019
in the section ‘Artifical Intelligence in Health’. The articles are listed in alphabetical
order of the first author's surname
|
Section
Artificial Intelligence in Health
|
|
• Albers DJ, Levine ME, Stuart A, Mamykina L, Gluckman B, Hripcsak G. Mechanistic
Machine Learning: How Data Assimilation Leverages Physiological Knowledge Using Bayesian
Inference to Forecast the Future, Infer the Present, and Phenotype. J Am Med Inform
Assoc 2018;25(10):1392-401.
• Lee J, Sun J, Wang F, Wang S, Jun CH, Jiang X. Privacy-Preserving Patient Similarity
Learning in a Federated Environment: Development and Analysis. JMIR Med Inform 2018;6(2):e20.
• Oktay O, Ferrante E, Kamnitsas K, Heinrich M, Bai W, Caballero J, Cook SA, de Marvao
A, Dawes T, O'Regan DP, Kainz B, Glocker B, Rueckert D. Anatomically Constrained Neural
Networks (ACNNs): Application to Cardiac Image Enhancement and Segmentation. IEEE
Trans Med Imaging 2018;37(2):384-95.
|
Conclusions and Outlook
In 2018, researchers published promising advances in the science of AI to address
some of the most significant challenges to its application in health. Two of the best
papers enhanced performance of learning methods by incorporating domain knowledge
to limit the learning space. Albers et al., re-introduced the technique of data assimilation, commonly used in aeronautical
applications, to the problem of predicting glucose values in type 2 diabetes. In this
study, they combined machine learning with an endocrine physiological model to make
accurate glucose level predictions, generated recommendations to produce desired outcomes,
and learned model parameters to represent phenotypes [3]. This approach allowed accurate prediction with sparse data, which can be common
in healthcare, and supported trust through explainability. The authors highlighted
the need for a multidisciplinary biomedical informatics pipeline to apply data assimilation
to the domain of medicine. Oktay et al., presented a novel strategy to train neural networks to learn underlying anatomy
from cardiac imaging data using a stacked autoencoder constrained by global anatomic
knowledge [11]. In this study, anatomic rather than physiologic knowledge was leveraged to demonstrate
excellent performance in image segmentation tasks using multimodal magnetic resonance
imaging (MRI) and ultrasound data sets. Importantly, this approach was able to overcome
the very common problems of motion artifact and a lack of internal consistency, which
limit many approaches to image interpretation. Further, although the methodology was
designed for segmentation, the algorithm showed good performance in pathological classification.
The third best paper addressed the challenging task of applying learning models in
data sets that are distributed across institutions. Lee et al., presented a novel privacy-preserving analytics platform for patient similarity learning
in a federated setting, through a multi-hash approach for context dependent cross-institution
patient representation, and incorporation of homomorphic encryption for privacy preservation
[8]. This work addressed the important problem of enabling privacy-preserving learning
in healthcare, where sufficient data to make inferences might be stored across a wide
variety of sites.
Five of the remaining selected papers described advances in medical imaging analysis
through novel application and enhancements of deep learning methods. Shi et al., applied an emerging technique of stacked deep polynomial networks to a multimodal
imaging-classification task [14]. These scientists used a staged approach to building machine learning algorithms,
applying stacked deep polynomial networks to MRI and positron emission tomography
(PET) images to fuse and learn feature representation to classify images from patients
with Alzheimer’s disease, normal controls, and mild cognitive impairment who subsequently
did or did not progress to Alzheimer’s disease. Their approach demonstrated excellent
accuracy in a challenging multi-class classification task for categories of a disease
that often progresses gradually and unpredictably. While MRI has become an indispensable
diagnostic tool, its slow acquisition process makes it expensive and less accessible.
To address this challenge, Schlemper et al., proposed a well-designed and well-executed framework for reconstructing dynamic sequences
of two-dimensional cardiac MRI images from aggressively under-sampled data using a
deep cascade of convolutional neural networks (CNNs) to accelerate the data acquisition
process [13]. Liu et al., focused on another difficult task in using MRI for diagnosis of brain diseases including
Alzheimer’s disease - that of automatic generation of clinically meaningful features.
They proposed a novel approach of identifying discriminative regions called landmarks,
followed by application of a CNN for patch-based deep feature learning [10]. Addressing challenges in other common imaging modalities, Gruezemacher et al., developed a three-dimensional (3D) adaptation of deep neural nets (DNNs) for lung
nodule detection from computed tomography (CT) images [6]; Hassan et al., presented a novel approach combining tensor-based segmentation, Delauday triangulation,
and deep learning models to provide complete 3D presentation of macula to support
automated diagnosis of various macula conditions [7].
Chang et al., addressed another important problem of how to train deep learning models across
distributed data sets as an alternative to data sharing [5]. They proposed three non-parallel heuristics (i.e., ensembling single institution
models, single weight transfer, and cyclic weight transfer) for training deep learning
models using distributed data and produced valuable insights through a set of very
well designed simulations run on three public data sets. Cyclic weight transfer resulted
in performance similar to centrally hosted data, even with low resolution images and
institutional variability in class distributions.
Successful development of deep learning models requires abundant labeled data, which
is particularly difficult to obtain in medical domain. Li et al., proposed an approach that leverages a type of multi-task predictive neural network
called generalized auxiliary-task augmented network (GATAN) to avoid overfitting [9], and they demonstrated improved performance when the learning algorithm has access
to limited labeled data.
AI methods for clinical concept and knowledge extraction from text is another important
topic, particularly on use of deep leaning methods to reduce the need for human-engineered
features and thus improve generalizability. Luo et al., proposed a CNN-based method called segment convolutional networks (seg-CNNs) that
used word embedding only for the extraction of entity relationships in clinical notes
with promising results [17].
Causal inference is an area that is receiving increasing attention, as the field moves
from initial proof-of-concept predictive models demonstrating the value of real world
data analysis to methods that can provide more actionable insights. Estimating the
heterogeneous effect of a treatment using observation data such as that found in electronic
health records (EHRs) is one of the most common and challenging use cases of causal
inference. Powers et al., compared and evaluated a number of well know methods for estimating heterogeneous
treatment effects through a preliminary study using the SPRINT data set [12], [18]. They identified causal boosting and causal multivariate adaptive regression splines
as two of the most promising methods and emphasized the need for much more effort
in this area.
Three of the selected papers covered topics in genomics, an area that has seen increasing
adoption of AI methods. Uppu and Krishna proposed a hybrid approach combining DNN
and random forest for the detection of multi-locus interactions among single nucleotide
polymorphisms [15]. Xiao et al., proposed a semi-supervised learning framework based on stacked sparse auto-encoder
for cancer prediction from RNA-seg data [16]. Belciug and Gorunescu proposed a new learning method based on extreme learning
machine to train a single layered feed forward network on microarray and mass spectrometry
data for cancer prediction [4].
