Introduction to Artificial Intelligence and Deep Learning with a Case Study in Analyzing Electronic Health Records for Drug Development
Harry Yang, Binbing Yu in Real-World Evidence in Drug Development and Evaluation, 2021
Deep learning is a re-branded name of an old class of methods, named neural networks, with new engineering tricks making the optimization of the networks faster and better. Neural networks are networks of artificial neurons mimicking how the human brain functions, where information is transmitted from one neuron to multiple other neurons after being processed in that neuron. The transmission process is one way. Figure 8.1 displays a simple feed-forward neural network with only one hidden layer. The predictors, x1 and x2, are fed into the input layer where each neuron represents one input variable. Then the data flow into the hidden layer in a weighted fashion and get transformed by a function f, called the activation function, which transforms the weighted input into an output. The output from each neuron in the hidden layer then flows into the output layer in a weighted fashion and, after being transformed by the activation function f, the response variable is predicted as . Usually the function f is a nonlinear function of the weighted sum of the inputs. If f is a linear function, the hidden layer is redundant and the neural network becomes a linear regression model. With more hidden layers, deep learning has the potential of modeling complicated interactions and nonlinear relationships between inputs and outputs.
Computational Neuroscience and Compartmental Modeling
Bahman Zohuri, Patrick J. McDaniel in Electrical Brain Stimulation for the Treatment of Neurological Disorders, 2019
In the 1980s, connectionism emerged as a prominent rival to classical computationalism. Connectionists draw inspiration from neurophysiology rather than logic and computer science. They employ computational models, neural networks that differ significantly from Turing-style models. A neural network is a collection of interconnected nodes. Nodes fall into three categories: input nodes, output nodes, and hidden nodes (which mediate between input and output nodes). Nodes have activation values, given by real numbers. One node can bear a weighted connection to another node, also given by a real number. Activations of input nodes are determined exogenously: these are the inputs to computation. Total input activation of a hidden or output node is a weighted sum of the activations of nodes feeding into it. Activation of a hidden or output node is a function of its total input activation; the particular function varies with the network. During neural network computation, waves of activation propagate from input nodes to output nodes, as determined by weighted connections between nodes.
The use of a Genetic Algorithm Neural Network (GANN) for Prognosis in Surgically Treated Nonsmall Cell Lung Cancer (NSCLC)
Raouf N.G. Naguib, Gajanan V. Sherbet in Artificial Neural Networks in Cancer Diagnosis, Prognosis, and Patient Management, 2001
Two important differences between the situation De Dombal et al. faced now mean that the need to retrain can be solved easily by methods such as the one implemented in this study. These are, firstly, neural network methods have the ability to learn by experience and evolve to optimal systems for the data with which they are used. Secondly, computing equipment is now generally available and inexpensive. Further, the cheapness and availability of powerful computing equipment means that, unlike the situation faced by De Dombal and his contemporaries, computational activity can take place at the site where the system is to be used. Such systems can then evolve to match particular institutions and doctors. The availability of computing equipment is so universal that there is no longer any need to attempt to provide a single set of system parameters that will match all doctors and all institutions. Indeed, a single set of system parameters (such as a set of likelihoods), which will be the final and complete answer, is unlikely given the wide variability of subjective clinical data. Furthermore, it may not be desirable, as a system dependent upon a single set of system parameters may not have the ability to adapt to changing circumstances [35]. This problem is the mathematical analogue of the problem of biological overspecialisation. An organism may be so well adapted for a given environ-ment and have so little capacity to change that it fails to survive should it face a somewhat different environment.
Advancing cervical cancer diagnosis and screening with spectroscopy and machine learning
Published in Expert Review of Molecular Diagnostics, 2023
Carlos A. Meza Ramirez, Michael Greenop, Yasser A. Almoshawah, Pierre L. Martin Hirsch, Ihtesham U. Rehman
Neural networks are computational models which can classify or organize datasets. Schematically, a neural network model consists of different processing units which are connected and run in parallel. These processing units aim to mimic a neuron from a biological brain where a series of decisions occur. A neural network may consist of an ‘n’ number of layers, each with an ‘n’ number of process units (Figure 2) [24,49,50]. The number of layers depends on the type of neural network; shallow or deep neural network. Whereas the number of processing units depends on the model input defined by the user. A shallow neural network may have between 5 and 10, but this number may differ [24,49,50]. The product of the mathematical process that occurs on each processing unit is classified by the activation function. In other words, the activation function is responsible for characterizing the data by classes, or groups [24,49,50]. Based on the author’s experience, in spectroscopy it is recommended to use ReLu, sigmoid, or ReLu activation functions, for binary or multiclass predictions, however, this is subject to the users’ criteria.
Current approaches for choosing feature selection and learning algorithms in quantitative structure–activity relationships (QSAR)
Published in Expert Opinion on Drug Discovery, 2018
ANN is the most widely used artificial intelligence technique in classification as well as regression-based QSAR [39–41]. ANN is often used for identification of most relevant features for the end point or response being modeled as well as for model development in QSAR/QSPR. Wikel and Dow in 1993 [42] first reported the application of ANN in feature selection; they have found that the back propagation neural network is useful and most efficient in the identification of most relevant descriptors in QSAR studies. Tetko et al. were the first to apply ANN as the variable selection technique for modeling of the activity in a set of HIV-1 reverse transcriptase inhibitors [43]. The era of the neural network techniques evolved with the computational technology, and in this era, the computer programs have been developed to learn from the data in a manner similar to the human brain. The ANN technique is mostly used when the situation is more complex such as a large number of descriptors, and possibly more complex descriptors. The ANN procedure to find the appropriate combination of descriptors is as follows:
Remaining challenges in predicting patient outcomes for diffuse large B-cell lymphoma
Published in Expert Review of Hematology, 2019
R. Andrew Harkins, Andres Chang, Sharvil P. Patel, Michelle J. Lee, Jordan S. Goldstein, Selin Merdan, Christopher R. Flowers, Jean L. Koff
With major advances in the fields of epidemiology, genomics, and clinical research, large amounts of heterogeneous data have become available in various health-care organizations. Therefore, there is a profound need for unified machine learning-based platforms incorporating vast amounts of mixed data types (e.g., imaging, histological, clinical, and genomic). The neural networks-based approaches, broadly described as deep learning, have been successfully implemented in areas such as image recognition, natural language processing, and robotics. Due to its ability to effectively leverage large data sets, the application of deep learning for precision genomic medicine is rapidly developing and has shown promise for the prediction of clinical outcomes with genomics [54,99]. Future efforts should also aim to integrate robust machine learning-based platforms into clinical use to improve the risk stratification of DLBCL patients in a manner that can eventually translate to more effective and personalized treatment strategies.
Related Knowledge Centers
- Artificial Intelligence
- Artificial Neural Network
- Axon
- Dendrite
- Habituation
- Synapse
- Neurotransmitter
- Neural Circuit
- Neuron
- Dendrodendritic Synapse