Leveraging Data Science for Global Health pp 199217  Cite as
Machine Learning for Clinical Predictive Analytics
 1.7k Downloads
Abstract
In this chapter, we provide a brief overview of applying machine learning techniques for clinical prediction tasks. We begin with a quick introduction to the concepts of machine learning, and outline some of the most common machine learning algorithms. Next, we demonstrate how to apply the algorithms with appropriate toolkits to conduct machine learning experiments for clinical prediction tasks. This chapter is composed of five sections. First, we will explain why machine learning techniques are helpful for researchers in solving clinical prediction problems (Sect. 12.1). Understanding the motivations behind machine learning approaches in healthcare are essential, since precision and accuracy are often critical in healthcare problems, and everything from diagnostic decisions to predictive clinical analytics could dramatically benefit from databased processes with improved efficiency and reliability. In the second section, we will introduce several important concepts in machine learning in a colloquial manner, such as learning scenarios, objective/target function, error and loss function and metrics, optimization and model validation, and finally a summary of model selection methods (Sect. 12.2). These topics will help us utilize machine learning algorithms in an appropriate way. Following that, we will introduce some popular machine learning algorithms for prediction problems (Sect. 12.3), for example, logistic regression, decision tree and support vector machine. Then, we will discuss some limitations and pitfalls of using the machine learning approach (Sect. 12.4). Lastly, we will provide case studies using real intensive care unit (ICU) data from a publicly available dataset, PhysioNet Challenge 2012, as well as the breast tumor data from Breast Cancer Wisconsin (Diagnostic) Database, and summarize what we have presented in this chapter (Sect. 12.5).
Keywords
Machine Learning Artificial Intelligence Healthcare Clinical Data Medical ComputingLearning Objectives

Understand the basics of machine learning techniques and the reasons behind why they are useful for solving clinical prediction problems.

Understand the intuition behind some machine learning models, including regression, decision trees, and support vector machines.

Understand how to apply these models to clinical prediction problems using publicly available datasets via case studies.
12.1 Why Machine Learning?
Machine learning is an interdisciplinary field which consists of computer science, mathematics, and statistics. It is also an approach toward building intelligent machines for artificial intelligence (AI). Different from rulebased symbolic AI, the idea of utilizing machine learning for AI is to learn from data (examples and experiences). Instead of explicitly programming handcrafted rules, we construct a model for prediction by feeding data into a machine learning algorithm, and the algorithm will learn an optimized function based on the data and the specific task. Such datadriven methodology is now the stateoftheart approach of various research domains, such as computer vision (Krizhevsky et al. 2012), natural language processing (NLP) (Yala et al. 2017), and speechtotext translation (Wu et al. 2016; Chung et al. 2018, 2019), for many complex realworld applications.
Due to the increased popularity of the electronic health record (EHR) system in recent years, massive quantities of healthcare data have been generated (Henry et al. 2016). Machine learning for healthcare therefore becomes an emerging applied domain. Recently, researchers and clinicians have started applying machine learning algorithms to solve the problems of clinical outcome prediction (Ghassemi et al. 2014), diagnosis (Gulshan et al. 2016; Esteva et al. 2017; Liu et al. 2017; Chung and Weng 2017; Nagpal et al. 2018), treatment and optimal decision making (Raghu et al. 2017; Weng et al. 2017; Komorowski et al. 2018) using data in different modalities, such as structured lab measurements (Pivovarov et al. 2015), claims data (DoshiVelez et al. 2014; Pivovarov et al. 2015; Choi et al. 2016), free texts (Pivovarov et al. 2015; Weng et al. 2018, 2019), images (Gulshan et al. 2016; Esteva et al. 2017; Bejnordi 2017; Chen et al. 2018), physiological signals (Lehman et al. 2018), and even crossmodal information (Hsu et al. 2018; Liu et al. 2019).
Instead of traditional adhoc healthcare data analytics, which usually requires expertintensive efforts for collecting data and designing limited handcrafted features, machine learningbased approaches help us recognize patterns inside the data and allow us to perform personalized clinical prediction with more generalizable models (Gehrmann et al. 2018). They help us maximize the utilization of massive but complex EHR data. In this chapter, we will focus on how to tackle clinical prediction problems using a machine learningbased approach.
12.2 General Concepts of Learning
12.2.1 Learning Scenario for Clinical Prediction

Define the outcome of your task

Consult with domain experts to identify important features/variables

Select an appropriate algorithm (or design a new machine learning algorithm) with a suitable parameter selection

Find an optimized model with a subset of data (training data) with the algorithm

Evaluate the model with another subset of data (testing data) with appropriate metrics

Deploy the prediction model on realworld data.
At the end of the chapter, we will show an exercise notebook that will help you go through the concepts mentioned above.
12.2.2 Machine Learning Scenarios
There are many machine learning scenarios, such as supervised learning, unsupervised learning, semisupervised learning, reinforcement learning, and transfer learning. We will only focus on the first two main categories, supervised learning and unsupervised learning. Both of the scenarios learn from the underlying data distribution, or to put it simply, find patterns inside data. The difference between them is that you have annotated data under the supervised scenario but only unlabelled data under unsupervised learning scenario.
12.2.2.1 Supervised Learning
Supervised learning is the most common scenario for practical machine learning tasks if the outcome is welldefined, for example, if you are predicting patient mortality, hospital length of stay, or drug response. In general, the supervised learning algorithm will try to learn how to build a classifier for predicting the outcome variable y given input x, which is a mapping function f where \(y=f(x)\). The classifier will be built by an algorithm along with a set of data \(\{x_1, ..., x_n\}\) with the corresponding outcome label \(\{y_1, ..., y_n\}\). Supervised learning can be categorized by two criteria, either by type of prediction or by type of model. First, it can be separated into regression or classification problems. For predicting continuous outcomes, using regression methods such as linear regression is suitable. For class prediction, classification algorithms such as logistic regression, naive Bayes, decision trees or support vector machines (SVM) (Cortes and Vapnik 1995) will be a better choice. For example, linear regression is suitable for children height prediction problem whereas SVM is better for binary mortality prediction.
Regarding the goal of the learning process, a discriminative model such as regression, trees and SVMs can learn the decision boundary within the data. However, a generative model like naive Bayes will learn the probability distributions of the data.
12.2.2.2 Unsupervised Learning
Without corresponding output variables (y), the unsupervised learning algorithms discover latent structures and patterns directly from the given unlabeled data \(\{x_1, ..., x_n\}\).
There is no ground truth in the unsupervised learning, therefore, the machine will only find associations or clusters inside the data. For example, we may discover hidden subtypes in a disease using an unsupervised approach (Ghassemi et al. 2014).
12.2.2.3 Other Scenario
Other scenarios such as reinforcement learning (RL) frame a decision making problem into a computer agent interaction with a dynamic environment (Silver et al. 2016), in which the agent attempts to reach the best reward based on feedback when it navigates the state and action space. Using a clinical scenario as an example, the agent (the RL algorithm) will try to improve the model parameters based on iteratively simulating the state (patient condition) and action (giving fluid or vasopressor for hypotension), obtain the feedback reward (mortality or not), and eventually converge to a model that may yield optimal decisions (Raghu et al. 2017).
12.2.3 Find the Best Function
Examples of commonlyused loss functions in machine learning
Task  Error type  Loss function  Note 

Regression  Meansquared error  \(\frac{1}{n} \sum _{i=1}^n (y_i  \hat{y_i})^2\)  Easy to learn but sensitive to outliers (MSE, L2 loss) 
Mean absolute error  \(\frac{1}{n} \sum _{i=1}^n y_i  \hat{y_i}\)  Robust to outliers but not differentiable (MAE, L1 loss)  
Classification  Cross entropy = Log loss  \( \frac{1}{n} \sum _{i=1}^n [y_i \log (\hat{y_i}) + (1y_i) \log (1\hat{y_i})] =  \frac{1}{n} \sum _{i=1}^n p_i \log q_i\)  Quantify the difference between two probability distributions 
Hinge loss  \(\frac{1}{n} \sum _{i=1}^n max(0, 1  y_i \hat{y_i})\)  For support vector machine  
KL divergence  \(D_{KL}(pq) = \sum _{i} p_i (\log \frac{p_i}{q_i})\)  Quantify the difference between two probability distributions 
Loss function defines the difference between the output of model y and the real data value \(\hat{y}\). Different machine learning algorithms may use different loss functions, for example, least squared error for linear regression, logistic loss for logistic regression, and hinge loss for SVM (Table 12.1). Cost function is the summation of loss functions of each training data point. Using loss functions, we can define the cost function to evaluate model performance. Through loss and cost functions, we can compute the performance of functions on the whole dataset.
In unsupervised learning setting, the algorithms have no real data value to compute the loss function. In such case, we can use the input itself as the output and compute the difference between input and output. For example, we use reconstruction loss for autoencoder, a kind of unsupervised learning algorithms, to evaluate whether the model can reconstruct the input from hidden states inside the model.
Commonlyused metrics in machine learning
Predicted  

True  False  
Actual  True  True positive (TP)  False negative (FN) Type II error  Recall = Sensitivity = \(\frac{\mathrm {TP}}{\mathrm{TP}\,+\,\mathrm{FN}}\) 
False  False positive (FP) Type I error  True negative (TN)  Specificity = \(\frac{\mathrm {TN}}{\mathrm{TN}\,+\,\mathrm{FP}}\)  
Precision = \(\frac{\mathrm {TP}}{\mathrm{TP}\,+\,\mathrm{FP}}\)  Accuracy = \(\frac{\mathrm{TP}\,+\,\mathrm{TN}}{\mathrm{TP}\,+\,\mathrm{TN}\,+\,\mathrm{FP}\,+\,\mathrm{FN}}\) F1 = \(\frac{\mathrm {2} \times \mathrm {Precision}\,\times \,\mathrm {Recall}}{{\mathrm{Precision}\,+\,\mathrm{Recall}}}\) 
12.2.4 Metrics
Choosing an appropriate numeric evaluation metric for optimization is crucial. Different evaluation metrics are applied to different scenarios and problems.
12.2.4.1 Supervised Learning
In classification problems, accuracy, precision/positive predictive value (PPV), recall/sensitivity, specificity, and the F1 score are usually used. We use a confusion matrix to show the relation between these metrics (Table 12.2).
12.2.4.2 Unsupervised Learning
Since there are no ground truth labels for unsupervised scenarios, evaluation metrics of unsupervised learning settings are relatively difficult to define and usually depend on the algorithms in question. For example, the CalinskiHarabaz index and silhouette coefficient have been used to evaluate kmeans clustering. Reconstruction error is used for autoencoder, a kind of neural network architecture for learning data representation.
12.2.5 Model Validation

Training set for model training. You will run the selected machine learning algorithm only on this subset.

Development (a.k.a. dev, validation) set, also called holdout, for parameter tuning and feature selection. This subset is only for optimization and model validation.

Testing set for evaluating model performance. We only apply the model for prediction here, but won’t change any content in the model at this moment.

It is better to have your training, dev and testing sets all from the same data distribution instead of having them too different (e.g. training/dev on male patients but testing on female patients), otherwise you may face the problem of overfitting, in which your model will fit the data too well in training or dev sets but find it difficult to generalize to the test data. In this situation, the trained model will not be able to be applied to other cases.

It is important to prevent using any data in the dev set or testing set for model training. Test data leakage, i.e. having part of testing data from training data, may cause the overfitting of the model to your test data and erroneously gives you a high performance but a bad model.
There is no consensus on the relative proportions of the three subsets. However, people usually allocate 20–30% of the whole dataset for their testing set. The proportion can be smaller if you have more data.
12.2.5.1 CrossValidation
The other commonly used approach for model validation is kfold cross validation (CV). The goal of kfold CV is to reduce the overfitting of the initial training set by further training several models with the same algorithm but with different training/dev set splitting.
Please keep in mind that the testing set is completely excluded from the process of CV. Only training and dev sets are involved in this process.
12.2.6 Diagnostics
After the first iteration of model training and evaluation, you may find that the trained model does not perform well on the unseen testing data. To address the issue of error in machine learning, we need to conduct some diagnostics regarding bias and variance in the model in order to achieve a model with low bias and low variance.
12.2.6.1 Bias and Variance
The bias of a model is the difference between the prediction model and the correct model for given data points. That is, it is the algorithm’s error rate on the training set. This is an underfitting problem, whereby the model can’t capture the trend of the data well due to an excessively simple model. One potential solution is to make the model more complex, which can be done by reducing regularization (Sect. 12.2.6.2), or configuring and adding more input features, for example, stacking more layers if you are using a deep learning approach. However, it is possible that the outcome of complex model is high variance.
The variance of a model is the variability of the model prediction for given data points. It is the model error rate difference between training and dev sets. Problems of high variance are usually related to the issue of overfitting. i.e. hard to generalize to unseen data. The possible solution is to simplify the model, such as using regularization, reducing the number of features, or add more training data. Yet the simpler model may also suffer from the issue of high bias.
High bias and high variance can happen simultaneously with very bad models. To achieve the optimal error rate, a.k.a. Bayes error rate, which is an unavoidable bias from the most optimized model, we need to do iterative experiments to find the optimal bias and variance tradeoff.
Finally, a good practice of investigating bias and variance is to plot the informative learning curve with training and validation errors. In Fig. 12.3 and Table 12.3 we demonstrate a few cases of diagnostics as examples.
12.2.6.2 Regularization
The goal of regularization is to prevent model overfitting and high variance. The most common regularization techniques include Least absolute shrinkage and selection operator (LASSO regression, L1regularization) (Tibshirani 1996), ridge regression (L2regression) (Hoerl and Kennard 1970), and elastic net regression (a linear combination of L1 and L2 regularization) (Zou and Hastie 2005).
The characteristic of high bias and high variance
Training error  Validation error  Approach  

High bias  High  Low  Increase complexity 
High variance  Low  High  Decrease complexity Add more data 
L1 and L2regularized logistic regression
Regularization  Equation 

L1 (LASSO)  \(\sum _{i=1}^m (y_i  \sum _{j=1}^n \beta _j x_{ij})^2 + \lambda \sum _{j=1}^n  \beta _j\) 
L2 (Ridge)  \(\sum _{i=1}^m (y_i  \sum _{j=1}^n \beta _j x_{ij})^2 + \lambda \sum _{j=1}^n \beta _j^2\) 
L1regularization is also a good technique for feature selection since it can “shrink” the coefficients of less important features to zero and remove them. In contrast, L2regularization just makes the coefficients smaller, but not to zero.
12.2.7 Error Analysis
It is an important practice to construct your first prediction pipeline as soon as possible and iteratively improve its performance by error analysis. Error analysis is a critical step to examine the performance between your model and the optimized one. To do the analysis, it is necessary to manually go through some erroneously predicted data from the dev set.
The error analysis can help you understand potential problems in the current algorithm setting. For example, the misclassified cases usually come from specific classes (e.g. patients with cardiovascular issues might get confused with those with renal problems since there are some shared pathological features between two organ systems) or inputs with specific conditions (Weng et al. 2017). Such misclassification can be prevented by changing to more complex model architecture (e.g. neural networks), or adding more features (e.g. combining word and conceptlevel features), in order to help distinguish the classes.
12.2.8 Ablation Analysis
Ablation analysis is a critical step for identifying important factors in the model. Once you obtain an ideal model, it is necessary to compare it with some simple but robust models, such as linear or logistic regression model. This step is also essential for research projects, since the readers of your work will want to know what factors and methods are related to the improvement of model performance. For example, the deep learning approach of clinical document deidentification outperforms traditional natural language processing approach. In the paper using neural network for deidentification (Dernoncourt et al. 2017), the authors demonstrate that the characterlevel token embedding technique had the greatest effect on model performance, and this became the critical factor of their study.
12.3 Learning Algorithms
In this section, we briefly introduce the concepts of some algorithm families that can be used in the clinical prediction tasks. For supervised learning, we will discuss linear models, treebased models and SVM. For unsupervised learning, we will discuss the concepts of clustering and dimensionality reduction algorithms. We will skip the neural network method in this chapter. Please refer to programming tutorial part 3 or a deep learning textbook for further information (Goodfellow et al. 2016).
12.3.1 Supervised Learning
12.3.1.1 Linear Models
Linear models are commonly used not only in machine learning but also in statistical analysis. They are widely adopted in the clinical world and can usually be provided as baseline models for clinical machine learning tasks. In this class of algorithms, we usually use linear regression for regression problems and logistic regression for classification problems.
The pros of linear models include their interpretability, less computational cost as well as less complexity comparing to other classical machine learning algorithms. The downside is their inferior performance. However, these are common tradeoff features in model selection. It is still worthwhile to start from this simple but powerful family of algorithms.
12.3.1.2 TreeBased Models

It looks across all possible thresholds across all possible features and picks the single feature split that best separates the data

The data is split on that feature at a specific threshold that yields the highest performance

It iteratively repeats the above two steps until reaching the maximal tree depth, or until all the leaves are pure.

Splitting criteria: by Gini index or entropy

Tree size: tree depth, tree pruning

Number of samples: minimal samples in a leaf, or minimal sample to split a node.
The biggest advantage of a decision tree is providing model interpretability and actionable decision. Since the tree is represented in a binary way, the trained tree model can be easily converted into a set of rules. For example, in their paper, Fonarow and colleagues utilized CART to create a series of clinical rules (Fonarow et al. 2005). However, decision trees may have high variance and yield an inferior performance.

Pick a random subset of features

Create a bootstrap sample of data (randomly resample the data)

Build a decision tree on this data

Iteratively perform the above steps until termination.
12.3.1.3 Support Vector Machine (SVM)
SVM is a very powerful family of machine learning algorithms (Cortes and Vapnik 1995). The goal of SVM is to attempt to find a hyperplane (e.g. a line in 2D, a plane in 3D, or a ndimension structure in a \(n+1\) dimensions space) to separate data points into two sides, and to maximize the minimal distance of the hyperplane from the sentinel data points (Fig. 12.4).
Regarding the optimization, we use hinge loss to train SVM. The pros of using SVM is its superior performance, yet the model’s inferior interpretability limits its applications in the healthcare domain.
12.3.2 Unsupervised Learning
In the previous section, we mentioned that the goal of unsupervised learning is to discover hidden patterns inside data. We can use clustering algorithms to aggregate data points into several clusters and investigate the characteristics of each cluster. We can also use dimensionality reduction algorithms to transform a highdimensional into a smallerdimensional vector space for further machine learning steps.
12.3.2.1 Clustering

Randomly initializing k points as the centroids of the k clusters

Assigning data points to the nearest centroid and forming clusters

Recomputing and updating centroids based on the mean value of data points in the cluster

Repeating step 2 and 3 until there is convergence.
The kmeans algorithm is guaranteed to converge to a final result. However, this converged state may be local optimum and therefore there is a need to experiment several times to explore the variability of results.
The obtained final k centroids, as well as the cluster labels of data points, can all serve as new features for further machine learning tasks, as shown in Sect. 9 of the “Applied Statistical Learning in Python” chapter. Regarding choosing the cluster number k, there are several techniques for k value validation. The most common methods include the elbow method, silhouette coefficient, or the CalinskiHarabaz index. However, it is very useful to decide k if you already have some clinical domain insights about potential cluster number.
12.3.2.2 Dimensionality Reduction
While dealing with clinical data, it is possible that you are faced with a very highdimensional but sparse dataset. Such characteristics may decrease the model performance even if you use machine algorithms such as SVM, random forest or even deep learning due to the risk of overfitting. A potential solution is to utilize dimensionality reduction algorithms to convert the dataset into lower dimensional vector space. Principal component analysis (PCA) is a method that finds the principal components of the data by transforming data points into a new coordinate system (Jolliffe et al. 2011). The first axis of the new coordinate system corresponds to the first principal component (PC1), which explains the most variance in the data and can serve as the most important feature of the dataset.
PCA is a linear algorithm and therefore it is hard to interpret the complex polynomial relationship between features. Also, PCA may not be able to represent similar data points of highdimensional data that are close together since the linear algorithm does not consider nonlinear manifolds.
The nonlinear dimensionality reduction algorithm, tDistributed Stochastic Neighbor Embedding (tSNE), becomes an alternative when we want to explore or visualize the highdimensional data (van der Maaten and Hinton 2008). tSNE considers probability distributions with random walk on neighborhood graphs on the curved manifold to find the patterns of data. Autoencoder is another dimensionality reduction algorithm based on a neural network architecture for learning data representation by minimizing the difference between the input and output of the network (Rumelhart et al. 1988; Hinton et al. 2006).
The dimensionality reduction algorithms are good at representing multidimensional data. Also, a smaller set of features learned from dimensionality reduction algorithms may not only reduce the complexity of the model, but also decrease model training time, as well as inference (classification/prediction) time.
12.4 Programming Exercise

Breast Cancer Wisconsin (Diagnostic) Database

Preprocessed ICU data from PhysioNet Challenge 2012 Database.

Learn how to use Google colab/Jupyter notebook

Learn how to build and diagnose machine learning models for clinical classification and clustering tasks.
In part 1, we will go through the basics of machine learning concepts through classification problems. In part 2, we will go deeper into unsupervised learning methods for clustering and visualization. In part 3, we will discuss deep neural networks. Please check the link of tutorials in the Appendix.
12.5 Pitfalls and Limitations
Machine learning is a powerful technique for healthcare research. From a technical and algorithmic perspective, there are many directions that we can undertake to improve methodology, such as generalizability, less supervision, multimodal training, or learning temporality and irregularity (Xiao et al. 2018).
However, there are some pitfalls and limitations about utilizing machine learning in healthcare that should be considered during model development (Chen et al. 2019). For example, model biases and fairness is a critical issue since the training data we use are usually noisy and biased (Caruana et al. 2015; Ghassemi et al. 2018). We still need human expertise to validate, interpret and adjust the models. Model interpretability is also an important topic from the aspects of (1) humanmachine collaboration and (2) building a humanlike intelligent machine for medicine (Girkar et al. 2018). Causality is usually not being addressed in most of the clinical machine learning research, yet it is a key feature of clinical decision making. We may need more complicated causal inference algorithms to inform clinical decisions.
We also need to think more about how to deploy the developed machine learning models into clinical workflow. How to utilize them to improve workflow (Horng et al. 2017; Chen et al. 2018), as well as integrate all information acquired by human and machine, to transform them into clinical actions and improve health outcomes are crucial for future clinicianmachine collaboration.
12.6 Conclusion
In summary, machine learning is an important and powerful technique for healthcare research. In this chapter, we have shown readers how to reframe a clinical problem into appropriate machine learning tasks, select and employ an algorithm for model training, perform model diagnostics and error analysis, as well as interpret model results. The concepts and tools described in this chapter aim to allow the reader to better understand how to conduct a machine learning project for clinical predictive analytics.
References
 AbuMostafa, Y. S., Lin, H. T., & MagdonIsmail, M. (2012). Learning from data: A short course. Amlbook.Google Scholar
 Bejnordi, B. E., Lin, J., Glass, B., Mullooly, M., Gierach, G. L., Sherman, M. E., et al. (2017) Deep learningbased assessment of tumorassociated stroma for diagnosing breast cancer in histopathology images. In 2017 IEEE 14th international symposium on biomedical imaging (ISBI 2017) (pp. 929–932). IEEE.Google Scholar
 Breiman, L. (2017). Classification and regression trees. Routledge.Google Scholar
 Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32.CrossRefGoogle Scholar
 Caruana, R., Lou, Y., Gehrke, J., Koch, P., Sturm, M., & Elhadad, N. (2015). Intelligible models for healthcare: Predicting pneumonia risk and hospital 30day readmission. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 1721–1730). ACM.Google Scholar
 Chen, P. C., Gadepalli, K., MacDonald, R., Liu, Y., Nagpal, K., Kohlberger, T., et al. (2018). Microscope 2.0: An augmented reality microscope with realtime artificial intelligence integration, Nature Medicine, 25(9), 14531457 arXiv:1812.00825. https://www.nature.com/articles/s4159101905397
 Chen, P. C., Liu, Y., & Peng, L. (2019). How to develop machine learning models for healthcare. Nature Materials, 18(5), 410.PubMedCrossRefGoogle Scholar
 Choi, Y., Chiu, C. Y.I., & Sontag, D. (2016). Learning lowdimensional representations of medical concepts. AMIA Summits on Translational Science Proceedings, 2016, 41.PubMedCentralGoogle Scholar
 Chung, Y.A., & Weng, W.H. (2017). Learning deep representations of medical images using Siamese CNNs with application to contentbased image retrieval. In Machine learning for health (ML4H) workshop at NIPS 2017.Google Scholar
 Chung, Y.A., Weng, W.H., Tong, S., & Glass, J. (2018). Unsupervised crossmodal alignment of speech and text embedding spaces. In Advances in neural information processing systems (pp. 7365–7375).Google Scholar
 Chung, Y.A., Weng, W.H., Tong, S., & Glass, J. (2019). Towards unsupervised speechtotext translation. In 2019 IEEE international conference on acoustics, speech and signal processing (ICASSP) (pp. 7170–7174). IEEE.Google Scholar
 Cortes, C., & Vapnik, V. (1995). Supportvector networks. Machine Learning, 20(3), 273–297.Google Scholar
 Dernoncourt, F., Lee, J. Y., Uzuner, O., & Szolovits, P. (2017). Deidentification of patient notes with recurrent neural networks. Journal of the American Medical Informatics Association, 24(3), 596–606.PubMedGoogle Scholar
 DoshiVelez, F., Ge, Y., & Kohane, I. (2014). Comorbidity clusters in autism spectrum disorders: An electronic health record timeseries analysis. Pediatrics, 133(1), e54–e63.PubMedPubMedCentralCrossRefGoogle Scholar
 Esteva, A., Kuprel, B., Novoa, R. A., Ko, J., Swetter, S. M., Blau, H. M., et al. (2017). Dermatologistlevel classification of skin cancer with deep neural networks. Nature, 542(7639), 115.PubMedPubMedCentralCrossRefGoogle Scholar
 Fonarow, G. C., Adams, K. F., Abraham, W. T., Yancy, C. W., Boscardin, W. J., Scientific Advisory Committee, A. D. H. E. R. E., et al. (2005). Risk stratification for inhospital mortality in acutely decompensated heart failure: Classification and regression tree analysis. JAMA, 293(5), 572–580.PubMedCrossRefGoogle Scholar
 Freund, Y., Schapire, R., & Abe, N. (1999). A short introduction to boosting. Journal Japanese Society for Artificial Intelligence, 14(771–780), 1612.Google Scholar
 Friedman, J. H. (2001). Greedy function approximation: A gradient boosting machine. In Annals of statistics (pp. 1189–1232).Google Scholar
 Gehrmann, S., Dernoncourt, F., Li, Y., Carlson, E. T., Wu, J. T., Welt, J., et al. (2018). Comparing deep learning and concept extraction based methods for patient phenotyping from clinical narratives. PloS One, 13(2), e0192360.PubMedPubMedCentralCrossRefGoogle Scholar
 Ghassemi, M., Naumann, T., DoshiVelez, F., Brimmer, N., Joshi, R., Rumshisky, A., et al. (2014). Unfolding physiological state: Mortality modelling in intensive care units. In Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 75–84). ACM.Google Scholar
 Ghassemi, M., Naumann, T., Schulam, P., Beam, A. L., & Ranganath, R. (2018). Opportunities in machine learning for healthcare. arXiv:1806.00388.
 Girkar, U. M., Uchimido, R., Lehman, L.W. H., Szolovits, P., Celi, L., & Weng, W.H. (2018). Predicting blood pressure response to fluid bolus therapy using attentionbased neural networks for clinical interpretability. In Machine learning for health (ML4H) workshop at NeurIPS 2018.Google Scholar
 Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learning. MIT Press.Google Scholar
 Gulshan, V., Peng, L., Coram, M., Stumpe, M. C., Derek, W., Narayanaswamy, A., et al. (2016). Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA, 316(22), 2402–2410.PubMedPubMedCentralCrossRefGoogle Scholar
 Henry, J., Pylypchuk, Y., Searcy, T., & Patel, V. (2016). Adoption of electronic health record systems among us nonfederal acute care hospitals: 2008–2015. ONC Data Brief, 35, 1–9.Google Scholar
 Hinton, G. E., & Salakhutdinov, R. R. (2006). Reducing the dimensionality of data with neural networks. Science, 313(5786), 504–507.PubMedPubMedCentralCrossRefGoogle Scholar
 Hoerl, A. E., & Kennard, R. W. (1970). Ridge regression: Biased estimation for nonorthogonal problems. Technometrics, 12(1), 55–67.CrossRefGoogle Scholar
 Horng, S., Sontag, D. A., Halpern, Y., Jernite, Y., Shapiro, N. I., & Nathanson, L. A. (2017). Creating an automated trigger for sepsis clinical decision support at emergency department triage using machine learning. PloS One, 12(4), e0174708.PubMedPubMedCentralCrossRefGoogle Scholar
 Hsu, T.M. H., Weng, W.H., Boag, W., McDermott, M., & Szolovits, P. (2018). Unsupervised multimodal representation learning across medical images and reports. In Machine learning for health (ML4H) workshop at NeurIPS 2018.Google Scholar
 Jolliffe, I. (2011). Principal component analysis. Springer.Google Scholar
 Komorowski, M., Celi, L. A., Badawi, O., Gordon, A. C., & Faisal, A. A. (2018). The artificial intelligence clinician learns optimal treatment strategies for sepsis in intensive care. Nature Medicine, 24(11), 1716.Google Scholar
 Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems (pp. 1097–1105).Google Scholar
 Lehman, E. P., Krishnan, R. G., Zhao, X., Mark, R. G., & Lehman, L.W. H. (2018). Representation learning approaches to detect false arrhythmia alarms from ECG dynamics. In Machine learning for healthcare conference (pp. 571–586).Google Scholar
 Liu, Y., Gadepalli, K., Norouzi, M., Dahl, G. E., Kohlberger, T., Boyko, A., et al. (2017). Detecting cancer metastases on gigapixel pathology images. arXiv:1703.02442.
 Liu, G., Hsu, T.M, H., McDermott, M., Boag, W., Weng, W.H., Szolovits, P., et al. (2019). Clinically accurate chest xray report generation. In Machine Learning for HealthcarearXiv:1904.02633.
 Nagpal, K., Foote, D., Liu, Y., Wulczyn, E., Tan, F., Olson, N., et al. (2019). Development and validation of a deep learning algorithm for improving gleason scoring of prostate cancer. NPJ digital medicine, 2(1), 110. arXiv:1811.06497. https://www.nature.com/articles/s4174601901122
 Pivovarov, R., Perotte, A. J., Grave, E., Angiolillo, J., Wiggins, C. H., & Elhadad, N. (2015). Learning probabilistic phenotypes from heterogeneous EHR data. Journal of Biomedical Informatics, 58, 156–165.PubMedCrossRefGoogle Scholar
 Raghu, A., Komorowski, M., Celi, L. A., Szolovits, P., & Ghassemi, M. (2017). Machine Learning for Healthcare: Continuous statespace models for optimal sepsis treatmenta deep reinforcement learning approach.Google Scholar
 Rumelhart, D. E., Hinton, G. E., Williams, R. J., et al. (1988). Learning representations by backpropagating errors. Cognitive Modeling, 5(3), 1.Google Scholar
 Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., Van Den Driessche, G., et al. (2016). Mastering the game of go with deep neural networks and tree search. Nature, 529(7587), 484.PubMedCrossRefGoogle Scholar
 Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society: Series B (Methodological), 58(1), 267–288.Google Scholar
 van der Maaten, L., & Hinton, G. (2008). Visualizing data using tSNE. Journal of Machine Learning Research, 9, 2579–2605.Google Scholar
 WeiWeng, H., & Szolovits, P. (2018). Mapping unparalleled clinical professional and consumer languages with embedding alignment. In 2018 KDD workshop on machine learning for medicine and healthcare.Google Scholar
 Weng, W.H., Chung, Y.A., & Szolovits, P. (2019). Unsupervised clinical language translation. In 25th ACM SIGKDD conference on knowledge discovery and data mining (KDD 2019).Google Scholar
 Weng, W.H., Gao, M., He, Z., Yan, S., & Szolovits, P. (2017). Representation and reinforcement learning for personalized glycemic control in septic patients. In Machine learning for health (ML4H) workshop at NIPS 2017.Google Scholar
 Weng, W.H., Wagholikar, K. B., McCray, A. T., Szolovits, P., & Chueh, H. C. (2017). Medical subdomain classification of clinical notes using a machine learningbased natural language processing approach. BMC Medical Informatics and Decision Making, 17(1), 155.PubMedPubMedCentralCrossRefGoogle Scholar
 Wu, Y., Schuster, M., Chen, Z., Le, Q. V., Norouzi, M., Macherey, W., et al. (2016). Google’s neural machine translation system: Bridging the gap between human and machine translation. arXiv:1609.08144.
 Xiao, C., Choi, E., & Sun, J. (2018). Opportunities and challenges in developing deep learning models using electronic health records data: A systematic review. Journal of the American Medical Informatics Association, 25(10), 1419–1428.PubMedPubMedCentralCrossRefGoogle Scholar
 Yala, A., Barzilay, R., Salama, L., Griffin, M., Sollender, G., Bardia, A., et al. (2017). Using machine learning to parse breast pathology reports. Breast Cancer Research and Treatment, 161(2), 203–211.PubMedCrossRefGoogle Scholar
 Zou, H., & Hastie, T. (2005) elasticnet: Elastic net regularization and variable selection. In R package version (p. 1).Google Scholar
Copyright information
Open Access This chapter is licensed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.
The images or other third party material in this chapter are included in the chapter's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the chapter's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.