Before reading this article, please click "Follow", which is convenient for you to join the discussion and sharing, and can bring you a different reading experience, thank you for your support
Wen | Mo Qingyan
Editor|Mo Qingyan
preface
Sample size estimation is an important part of clinical trials because the latter is the cornerstone of ensuring safety and efficacy, and a representative sample of sufficient size can provide insights into a given population, but collecting large amounts of data can be costly and time-consuming.
Each clinical trial must be carefully organized by developing a protocol that outlines the objectives of the study, primary and secondary endpoints, data collection methods, sample selection criteria, data processing procedures, statistical methods and hypotheses, and most importantly, a scientifically sound sample size.
Depending on the study design, outcome type, and hypothesis testing specified by the investigator, the determination of sample size can vary widely, and estimates of appropriate sample size are based on, given statistical assumptions and several study design parameters.
Theoretically, these factors include, meaningful, least detectable differences in the measurement, the level of statistical power required to estimate variability, and the level of significance.
It is imperative to achieve an optimal balance between the number of subjects, under- or too many subjects in theory, and insufficient statistical power due to small sample sizes may result in the inability to detect true differences.
As a result, significant differences between study groups are not statistically significant, and using too large sample sizes may be considered unethical, result in waste of resources, and may hinder the feasibility of a given study.
In this case, computational alternatives can be found to increase the sample size and thus reduce human exposure, and data augmentation is a method used to extend the sample by generating a modified copy of a given data set by using pre-existing data.
This involves minor modifications to datasets or leveraging deep learning techniques to generate new data instances, and the potential of AI as a means to achieve sustainable and enhanced drug development has been recognized.
However, the process of generating useful insights from accessible data requires leveraging comprehensive AI models, just as they must be built with the right datasets, and trained to effectively accelerate and streamline every step of drug research.
Recently, the Allen Turing Institute launched a project whose main goal is to study machine learning and the potential impact of artificial intelligence on the planning, conduct, and interpretation of randomized clinical trials.
By augmenting human expertise and optimizing data utilization, AI is able to predict the likelihood of trial or site failure, as well as the prognosis of clinical patients.
AI can also be used to analyze health records and effectively inform medical staff and patients of available trial opportunities, in addition to streamlining entry criteria could improve accessibility for potential participants.
In this regard, autoencoders have been identified as a very effective and valuable technique for generating synthetic or artificial data from real-world data, and autoencoders are a variant of artificial neural networks.
Used to obtain proficiency coding of unannotated data, the autoencoder acquires knowledge through two different functions, modifies the encoding function of the input data, and reconstructs the decoding function of the input data from the encoding representation.
A variational autoencoder is a generative model that runs under a probabilistic framework with neural networks as part of its broader architecture.
The aim of this study is to introduce a new idea of data enhancement in clinical trials by using variational autoencoders to reduce the required sample size.
To accomplish this task, several forms of variational autoencoders were explored and used for the generation of virtual populations.
The subjects generated by variational autoencoders, appropriately set up in the form of equivalence studies, analyzed are to tune the variational autoencoder system by selecting the most appropriate hyperparameters.
Materials and methods
Artificial intelligence refers to the replication of human intelligence in machines and computer systems, and artificial intelligence involves the development of intelligent machines that have the ability to perform tasks that are equal to or more than humans.
The process, which involves collecting data, developing guidelines for use, drawing approximations or conclusions, and self-correction, has attracted a lot of interest in AI and machine learning methods in several fields, including pharmaceutical science.
Deep learning involves learning, and continuous presentation layers that are increasingly correlated with data, known as models of neural networks, are organized in layers that are stacked on top of each other, and are used in deep learning to learn these hierarchical representations.
The most cutting-edge deep learning techniques take advantage of the latest advances in neural networks, which generally have better prediction and generalization skills than other machine learning techniques, drug discovery and repurposing, and are two areas of pharmaceutical research that deep learning has aroused interest and gained recognition.
Theoretically, in clinical trials, the sample size is clearly stated in the study protocol, so estimation before the start of the study, the estimation of the sample size is a complex process that relies on several parameters of the trial.
The most important of these are the endpoint measured, the scale of the measurement of the endpoint, the variability of the endpoint, the need for a large sample size, the latter resulting in the inclusion of many human participants rather than increased cost, long study duration, high probability of dropping out, etc.
In recent years, neural networks as the most advanced machines, learning models have become more and more popular, there are a variety of neural network architectures, each architecture is best suited to solve different types of problems, fully connected neural networks are widely regarded as, the most common type of neural network.
The basic components of a neural network are neurons, also known as nodes, and the vertical arrangement of many neurons creates a layer, the initial layer of the neural network is often called the input layer, and the final layer is designated as the output layer.
The number of nodes in the input layer corresponds to the number of features in the model, while the number of nodes in the output layer corresponds to the desired output, the middle layer in the neural network architecture is often called the hidden layer, the number of hidden layers, may vary depending on the complexity of the problem at hand.
In some cases, the use of large sample sizes may expose more participants to potential risks associated with experimental interventions, and participation in clinical trials can extend their duration.
This increases the chance that external factors influence the results and may affect the validity of the study, and the likelihood of loss to follow-up and drop-out during the trial is higher due to the larger number of participants.
Over the past few years, efforts have been made to conduct in silico clinical trials and/or actually increase the sample size in the face of all the shortcomings mentioned earlier, data augmentation techniques, which involve generating additional data points based on existing data, which effectively increase the sample size and improve the robustness of the analysis.
To accomplish this task, a normal distribution was used to generate a dataset for two groups of volunteers while simulating the conditions of an equivalent clinical trial, and the most important factor affecting the outcome of a clinical trial was the average difference between the two interventions.
Ideally, the performance of the generated dataset is better than that observed from the subsampled population, in other words, when the resulting dataset does exist.
And when the original dataset also shows equivalence, the resulting dataset is best shown non-equivalence, and any deviation from the original dataset indicates a defect in the analytical dataset.
To investigate the efficiency of the variational autoencoder method, the performance of the data set generated by the variational autoencoder was compared with the performance of the original data set, as well as the performance of the subsampled dataset, and in all cases, the performance of the data generated by the variational autoencoder was better than any subsampling group.
And with the original data, it is worth noting that for high variability, the performance of the variational autoencoder method is even better than the original data set.
Variational autoencoders have been successfully applied in different scientific fields, and in a study conditional variational autoencoders are used for intrusion detection in IoT networks, the latter importance is very important.
Because the economic importance of intrusion detection systems continues to expand, making them vulnerable to future intrusion attacks, the researchers used a conditional variational autoencoder with a unique architecture.
This variational autoencoder, which adds intrusion tags to the decoder layer, has many advantages compared with other well-known classifiers.
Because the inclusion or exclusion of certain options in the selection set results in inaccurate estimation parameters, the authors develop a generalized extrema model for linking value-added evaluation methods to selection modeling.
This study is considered to be the first study of various mathematical properties related to variational autoencoder models, which demonstrate proficiency in data clustering.
And it is observed that it is particularly effective in generating images or management of, a particular graphic structure, and creating images with low resolution requirements.
The performance of this experiment, compared with that of other established anomaly detection algorithms, shows that the developed model exhibits excellent performance in various situations and is considered by all participants.
Variational autoencoders are proposed as a data augmentation method in clinical trials, i.e. as a way to reduce the required sample size.
Variational autoencoders can be thought of as an extension of traditional autoencoders, compared to autoencoders, which employ encoder networks.
Map the input data to a multivariate normal distribution rather than a fixed point, in other words, variational autoencoders strive to establish correlations between input data and probability distributions across the potential space.
Without altering the real results of the study, it is worth noting that variational autoencoders are used even for data that exhibit high variability.
It is also possible to reduce the demand for samples to less than half of the sample size that trials usually require, and a significant reduction in sample size can accelerate clinical trials and significantly reduce costs.
And certainly reducing human exposure, to the best of our knowledge, this study represents the first attempt to use autoencoders within a clinical research area aimed at reducing the necessary sample size.
However, the trial is often used in clinical trials, and other clinical designs and statistical hypotheses must be explored at the request of regulatory agencies and interpreted in detail from a regulatory perspective.
Also with regard to model architecture, it may be worthwhile considering computational efficiency to look at more complex models that contain additional hidden layers, or more neurons per layer.
It is worth exploring other activation functions that can be applied to the hidden and output layers, and finally as with any calculation method.
It must be applied to real clinical data before being adopted in practice, and it should be emphasized that integrating real data into simulation studies is essential to improve the accuracy, reliability, and applicability of simulations.
Making it a valuable tool for understanding real-world phenomena, making predictions and informing decision-making processes, applying the introduced variational autoencoder program to a comparative analysis of actual clinical data will help determine its applicability and efficiency.
conclusion
The purpose of this study is to introduce the use of neural networks, in particular variational autoencoders, as a tool to virtually increase the sample size in clinical studies and thus reduce the number of actual participants required, this study first developed the most appropriate architecture for variational autoencoders and adjusted hyperparameters.
and various scenarios that can occur in practice, including measuring several levels of variability at the endpoint, different average performance between comparison groups, and different sizes of subsampling groups.
The efficiency of the data generated by the variational autoencoder is then compared with the efficiency of the original data, as well as the efficiency of the subsampled data, and in all cases, the data generated using the variational autoencoder improves the statistical power of the study.
Overall, variational autoencoders, combined with Monte Carlo simulation clinical trials, demonstrated desirable performance, thereby reducing human exposure in clinical studies.
and significantly reducing the cost and time to trial completion, as far as we currently understand, this study represents a new effort to use autoencoders and neural networks in clinical research, specifically aimed at reducing the necessary sample size.
After reading this article, please follow + comment, so that the remaining subsequent events have new developments that can be quickly followed