Dave's PCF WIP: Terms

Term #1,711

Edit Term

Artificial Neural Network and Machine Learning and deeplearning etc

View List Terms Delete

Claim Term

Application

App Docket

Created Date

Full Desc

Artificial Neural Networks

Artificial Intelligence can utilize, for example, neural networks. Artificial neural networks (ANNs) can learn tasks based on examples, e.g. without task specific programming. ANNs can be based on a group of connected units or nodes, i.e. artificial neurons. Each connection between artificial neurons can generate a signal to be transmitted to another artificial neuron. One or more artificial neurons can receive the signal and process it and, for example, use it to initiate a task. The signal at a connection between artificial neurons can be a number, and the outputs can be calculated by various functions or algorithms, e.g. non-linear functions. Artificial neurons can have a weight assigned to them, which can amplify or de-emphasize their signal. Artificial neurons can be organized in layers, with different layers performing different kind of transformations. Artificial neural networks can utilize various techniques, processes and/or algorithms, e.g. backpropagation, parallel distributed processing, max-pooling, Hebbian learning, long term potentiation, support vector machines, and linear classifiers. ANNs can include recurrent neural networks and deep feedforward neural networks. ANNs can perform functions such as, for example, pattern recognition and machine learning. Components of ANNs can include neurons, connections and weights, propagation functions, and learning rules. Neurons can include an activation component, a threshold component, an activation function, and an output function. ANNs can define mathematical and other functions. ANNs can use predefined functions, e.g. hyperbolic tangent function, sigmoid function, softmax function or rectifier function.

ANNs can be used for learning. Learning can comprise using a number of observations to find a function which solves a predetermined or desired task in an optimal sense, e.g. an optimal outcome. Learning can be supervised learning, unsupervised learning and reinforcement learning. Supervised learning can use one or more sets of example pairs and the goal can be, for example, to find a function in an allowed class of functions that matches the examples. Pattern recognition, classification, and regression can be part of supervised learning. Supervised learning can use informational fuzzy networks, random forests, nearest neighbor algorithms, logistic model tree, and other algorithms. Supervised learning can use statistical classification, including, for example, decision trees, Bayesian networks, and/or linear classifiers.

In unsupervised learning, a set of data can be provided for example along with a cost function to be minimized, which can be a function of the data and the network output. The cost function can be dependent on the task and the properties of the parameters and observed variables or data. Unsupervised learning can be applied, for example, to pattern recognition, classification, and regression, general estimation problems, clustering, the estimation of statistical distributions, compression and filtering. Unsupervised learning can use one or more ANNs, expectation-maximization algorithms, data clustering, and the like. Association rule learning can use a priori algorithms, eclat algorithms, FP-growth algorithms, hierarchical clustering (e.g., single-linkage clustering and conceptual clustering), partitional clustering (e.g., K-means algorithm, fuzzy clustering), reinforcement learning (e.g., Monte Carlo method, Q-learning, temporal difference learning, and combinations thereof.

In reinforcement learning, data can be generated by an agent's interactions with one or more objects, e.g. a surgeon interacting with a patient. The agent, e.g. the surgeon, can perform an action, and the environment, e.g. a target tissue or a surgical site, can generate one or more observations and, for example, a cost according to some dynamics or parameters, e.g. a tissue removal or an infection risk. The objective can be to discover a treatment, treatment algorithm, treatment modification that can reduce or minimize a measure of the cost, e.g. an infection risk, a patient reported outcome measurement, a functional result. The parameters and dynamics of the environment, e.g. a surgical site, can be unknown, but can be estimated. The environment, e.g. a target tissue or a surgical site, can be modeled as a Markov decision process and actions, with possible probability distributions, e.g. a cost distribution, an observation distribution, and one or more transitions, and a policy or algorithm or solution can be defined as a conditional distribution over actions given one or more observations. Dynamic programming can be coupled with ANNs and applied to multi-dimensional nonlinear problems.

Learning can utilize one or more cost functions, e.g. the cost being an excellent or a poor clinical outcome. The cost function can yield information of how far a particular solution, e.g. a clinical treatment, treatment sequence or treatment algorithm or surgical technique, is from an optimal outcome, e.g. an excellent score in a patient reported outcome measure. ANNs can find the solution, e.g. a clinical treatment, treatment sequence or treatment algorithm or surgical technique, that yields the lowest cost, e.g. distance or amount away from an optimal outcome or excellent score in a patient reported outcome. The cost can be a function of the observations. The cost can be described as a statistic. A cost can be the mean squared error, which can try to minimize the average squared error between the network's output and one or more target values over example pair(s). A cost function can be selected or predetermined for a particular problem set, e.g. a clinical problem set or clinical observation data, e.g. pre-operative, intra-operative or post-operative data. AI can find and develop one or more optimal cost functions for a set of observational data and AI can refine the cost function as the size of the observational data set increases. (See US Patent 11,278,413 Para. 58-63).

Machine Learning

Machine learning can comprise supervised learning, semi-supervised learning, active learning, reinforcement learning or unsupervised learning. With supervised learning, the computer can receive example inputs and desired outputs, which can be provided from a database or using a learning tool; the objective is to learn one or more rules that map the inputs to the outputs. Semi-supervised learning can be different in that the computer can be given an incomplete training example input, optionally with some desired outputs missing. With active learning, the computer can only obtain training inputs for a limited set of examples, and the computer can optimize the choice of inputs to acquire labels for. With reinforcement learning, training data, e.g. inputs and desired outputs, can be given only as feedback to the program's actions in a dynamic environment, such as guiding a surgery. With unsupervised learning, no training input and/or output data are provided, leaving the computer and computer processor on its own to find structure in the inputs.

Machine learning can use processes such as, for example, classification, regression, clustering, density estimation, dimensionality reduction, and topic modeling. With classification, inputs can be divided into two or more classes, and, for example, the learning system can produce a model that assigns unseen inputs to one or more of these classes. Data can be classified, for example, into “excellent”, “good”, “acceptable” or “poor” outcome, e.g. clinical outcome. Numeric values or ranges of numeric values can be assigned to different classes, for example numeric values or ranges of numeric values from a patient reported outcome measurement, from a clinical reporting system, and/or from one or more electronic measurements. With regression, outputs can be continuous rather than discrete. Regression can be used, for example, when patient outcomes are continuous. With clustering, inputs can be divided into groups. The groups cannot be known beforehand; thus, with clustering learning can be unsupervised. With density estimation, the distribution of inputs in a given space or sample can be determined. With dimensionality reduction, inputs can be simplified by mapping them into a lower-dimensional space. With topic modeling, a machine learning system can be given a list of human language documents and can be tasked to find out which documents cover similar topics.

Developmental learning can include robotic learning, which can generate its own learning situations to acquire new skills through autonomous self-exploration and interaction, for example, with human teachers.

A goal of machine learning or deep learning can be to generalize from the experience. Generalization can be the ability of a learning machine or system to perform accurately on new, unseen inputs after having been trained on a training data set. Training examples can come from an unknown probability distribution and the learning machine or system can be tasked to build an input and output model that enables it to produce sufficiently accurate predictions with new inputs. Bounds or limits, e.g. probabilitistic bounds, on the performance, accuracy and reproducibility of machine learning can be determined. When machine learning is used for solving clinical problems, e.g. outcome prediction or treatment planning or modification, the bounds or limits of performance, accuracy and/or reproducibility of the machine learning system or learning machine can influence and/or determine the performance, accuracy, and/or reproducibility of the clinical application, e.g. outcome prediction or treatment planning or modification. Accuracy can include the assessment of true positives, true negatives, false positives and false negatives. Reproducibility can be precision. Performance of the machine learning system and/or learning machine can include other statistical measures known in the art for assessing the performance of a clinical system.

Learning systems including machine learning can use decision tree learning, association rule learning, artificial neural networks (ANNs), deep learning, inductive logic programming, support vector machines, clustering, Bayesian networks, reinforcement learning, representation learning, similarity and metric learning, sparse dictionary learning, biologic or genetic algorithms, rule based machine learning and learning classifier systems.

With decision tree learning, a decision tree can be used as a predictive model, which can map observations about one or more parameters to conclusions about the parameter's target value. With association learning, relations between variables or parameters can be identified in large databases. With ANNs, computations can be structured through an interconnected group of artificial neurons, processing information using a connected approach. ANNs can be non-linear data modeling took, using various statistical methods and approaches known in the art. Deep learning can employ multiple layers in an artificial neural network. Inductive logic programming (ILP) can utilize logic programming for rule learning, e.g. using a uniform representation for input examples, background knowledge, and hypotheses. Support vector machines (SVMs) can be a set of supervised learning methods used for classification and/or regression. A given a set of training examples can be marked as belonging to first category or a second category; an SVM training machine can build a model predicting whether a new input falls into the first or the second category.

Clustering can be the assignment of a set of observations into subsets, where data in each subset have similarities with regard to one or more parameters while data in different subsets can be dissimilar with regard to one or more parameters. Clustering techniques can provide information on similarity or dissimilarity, for example reflected in a similarity metric, a measurement of internal compactness or separation between different clusters. A Bayesian network can be a graphical model representing, for example, random variables and their conditional independencies. This can be shown in a directed acyclic graph. A Bayesian network can represent the probabilistic relationships between diseases and symptoms. A Bayesian network can be used to compute the probability(ies) of the presence of one or more disease.

With reinforcement learning, input and output pairs can never be presented; reinforcement learning can map a state, e.g. a clinical state of a patient, and can develop predictions, actions or treatment the system can make. With representation learning algorithms input information can be preserved but transformed to make it more useful, e.g. as a pre-processing step, prior to performing classification or predictions, allowing reconstruction of the inputs coming from the unknown data generating distribution.

Deep learning can utilize multiple levels of representation, e.g. in an ANN. Higher-level, e.g. more abstract, parameters or data can be defined as generating lower-level parameters or data. With similarity learning, the learning system or machine can be given pairs of data that are considered similar and pairs of less similar data. It can then learn a similarity function or a distance metric function that can predict if new objects are similar.

Rule-based machine can be identification and utilization of a set of relational rules that can represent the knowledge captured by the learning system. Learning classifier systems (LCS) can be a family of rule based machine learning algorithms or systems which can combine a discovery component with a learning component.

The accuracy of classification machine learning models can be evaluated using accuracy estimation techniques and statistical techniques and methods testing the accuracy, sensitivity, specificity, fake positive and false negative rates. Other statistical methods such as Receiver Operating Characteristic (ROC) and associated Area under the Curve (AUC) as well as Total Operating Characteristic (TOC) can be used. (See US Patent 11,278,413 Para. 64-75).

Deep Learning

Deep learning can include machine learning algorithms which can use multiple layers of non-linear processing units or elements. Each layer can use the output from a higher layer as input. Deep learning systems can work in a supervised setting, e.g. using one more classification systems. Deep learning systems can also work in an unsupervised setting, e.g. in the context of texture analysis or pattern recognition. Deep learning systems can learn multiple levels of representations that correspond to different levels of abstraction. The different levels can form an order or a hierarchy of concepts. The different layers of a deep learning system can reside in different layers of an artificial neural network, i.e. a deep neural network. They can include hidden layers in an ANN. Deep learning systems and deep ANNs can utilize Boltzmann machines. With deep learning systems, layers can correspond to layers of abstraction, e.g. across a deep neural network. Varying numbers of layers and layer sizes can provide different degrees of abstraction. Higher level, more complex concepts can be learned from lower level layers.

Deep neural networks (DNNs) can be one or more ANNs with multiple hidden layers between the input and output layers. DNNs can model complex non-linear relationships. DNNs can generate models where the object is expressed as a layered composition. DNNs can be feedforward networks in which data flows from the input layer to the output layer without looping back, DNNs can be recurrent neural networks or convolutional deep neural networks.

Deep learning algorithms can be applied to unsupervised learning tasks. This is an important benefit because unlabeled data can more abundant than labeled data. For example, in a clinical environment, a deep learning system with a multi-layered ANN can initially be trained using a classification of outcomes in a supervised fashion. As the data grow, the system can optionally learn in an unsupervised manner, for example by utilizing pattern recognition across large clinical datasets, which can include pre-operative, intra-operative and post-operative data. (See US Patent 11,278,413 Para. 76-79).

Classification

Classification can be a process of creating categories, in which data or objects can be recognized, differentiated or understood. A classification system can be an approach of accomplishing classification. Classification can be performed using mathematical classification, statistical classification, classification theorems, e.g. in mathematics, and attribute value systems. Classifications can be alphanumeric. Classifications can be single or multi-dimensional. Classifications can be single or multi-layered. Classifications can be color coded. An ANN can use a single classification system, e.g. in supervised learning. An ANN can use multiple classification systems. When multiple classification systems are used, they can reside in different layers of a DNN or deep learning system.

Classification can be the problem of identifying to which of a set of categories or sub-populations a new observation belongs; this can be determined, for example, using a training data set with observations whose category membership is known. For example, a diagnosis can be assigned to a patient as a category which can be described by measured data or characteristics such a heart rate, blood pressure, presence of absence of symptoms or combinations of symptoms. Classification can be a pattern recognition.

Individual observations or data can be divided into a set of quantifiable properties. These properties may be categorical, e.g. “a”, “b”, “c”, “d” etc. or ordinal, e.g. “excellent”, “very good”, “good”, “acceptable”, “average”, or “poor”. They can be integer or real valued. Observations or data can also be classified using similarity or distance functions, e.g. based on earlier observations or data. An algorithm that implements classification can be a classifier, A classifier can sometimes also be a mathematical function, e.g. implemented by a classification algorithm, that can map input data to a category.

Data can be classified, for example, into “excellent”, “good”, “acceptable” or “poor” outcome, e.g. one or more clinical outcomes or clinical outcome variables. Numeric values or ranges of numeric values can be assigned to different classes, for example numeric values or ranges of numeric values from a patient reported outcome measurement, from a clinical reporting system, and/or from one or more electronic measurements.

Discriminative vs. Generative Models and Networks

In machine learning, discriminative models can be distinguished from generative models.

1. Discriminative models are trained to learn the boundaries between classes. They model the conditional probability of a target variable Y (class), given an observation x: P(Y|X=x) (“probability of Y given X=x”). Discriminative models describe the probability for classifying a given example x into a classy E Y. Discriminative models include, for example, logistic regression, conditional random fields, support vector machines, neural networks, random forests, or perceptrons.

Generative models model the distribution of individual classes. They can generate data and provide a statistical model of the joint probability distribution on X×Y, P(X,Y)=P(X|Y)*P(Y), for an observable variable X and a target variable Y. Generative models include, for example, naïve Bayes models and Bayes networks, hidden Markov models, Boltzmann machines, variational autoencoders or generative adversarial networks (GAN).

In some embodiments, the computer system can use a trained artificial neural network (ANN) to determine the treatment plan. The ANN can implement a discriminative model. A discriminative model can be trained to classify the input data, i.e. the preoperative and/or intraoperative data and/or postoperative data, into different classes, wherein each class can represent a different treatment plan.

In some embodiments, the ANN can implement a generative model. Instead of assigning preoperative and/or intraoperative input data and/or postoperative data to an existing class, a generative model is trained to generate the treatment plan steps based on the input data.

In some embodiments, a generative and a discriminative network model can be combined into a generative adversarial network (GAN) to generate a treatment plan. Using a training data set of existing recorded treatment plans for a number of preoperative and/or intraoperative input data and/or postoperative data sets, in this situation, the generative network can be trained to generate a preferred treatment plan from the preoperative and/or intraoperative input data. The discriminative network can be trained to evaluate the generated treatment plan and to distinguish the generated treatment plan from the actual treatment plan of the training case. Thus, the discriminative network can force the generative network to improve its results. (See US Patent 11,278,413 Para. 80-91).