Book review: Engineering life-enhancing solutions with machine learning
By Shalin Hai-Jew, Kansas State University
Machine Learning Algorithms for Engineering Applications: Future Trends and Research Directions
Prasenjit Chatterjee, Parmanand Astya, Sudeshna Chakraborty, and Pooja, Editors
Nova Science Publishing
2022
208 pp.
Machine learning (ML) is a sub-element of artificial intelligence (AI), informally defined as synthetic intelligence inspired by biological human / animal / plant sources. ML expands what is knowable given sufficient and relevant information, from which data models are made; statistical and mathematical analyses are applied to understand potentially relevant data patterns.
Readers of the text are advised to know “R, CUDA, Mahout, Pytorch, TensorFlow, and CoLab” to work the examples (Chatterjee, Astya, Chakraborty, & Pooja, 2022, p. x). While programming is not apparently required, that knowledge cannot hurt. ML tools may run on local machines or remote ones. Machine learning platforms often come with “extensive libraries, a user-friendly GUI, and support for popular programming languages” (Chatterjee, Astya, Chakraborty, & Pooja, 2022, pp. x-xi).
There are three general paradigms of machine learning: (1) supervised machine learning (with labeled data), (2) unsupervised machine learning (without labeled data), and (3) reinforcement learning (a method that does not need labeled data, which applies dynamic programming applying an inexact Markov decision process to both explore and exploit data to optimize targeted learning). More on the approaches and related algorithms follow later on. This is done without dedicated or specific programming, but there is logic in how to set up ML for particular datasets and explorations. Others in their ML typography include ensemble learning (combining multiple AI approaches) and various types of artificial neural networks.
Figure 1. Globular
Parveen Mor and Sonia Chhabra’s “Tools for Machine Learning” (Ch. 1) describe machine learning as a “technique,” with ways for computers to find data patterns and “learn” while “not being expressly programmed” (p. 1). This team offers advice to those getting started with machine learning:
Common machine learning languages include Python, C++, and R, among others. Machine learning applications include NumPy, Scikit-learn, NLTK (Natural Language Toolkit in Python), and TensorFlow, likely all recognizable names.
Supervised machine learning takes as input labeled data, so the intelligent agents examine information based on historical knowledge. Unsupervised machine learning takes as inputs unlabeled data, but these approaches enable various clustering based on observed associations. Reinforcement machine learning also works without labeled data, and the methods in this paradigm work flexibly to maximize targeted learning towards a “cumulative reward” of outcomes. The reinforcement approaches enable learning from the available data to optimize desired outcomes, using a Markov decision process (without defining an exact mathematical model, often given the large sizes of Markov decision processes). They write of ML that it requires “deep neural networks and machine learning methods to assist programming” (Mor & Chhabra, 2022, p. 9).
It is encouraging to see quite a number of gateways to bridge to machine learning. There are numerous opportunities for experimentation, with public datasets, with labeled and unlabeled data.
What would enhance this chapter would be some more real-world examples of ML analytics. It would help to have some examples of how to approach a dataset and set it up for ML to explore particular dimensions of that data. In other words: How exploitable are large datasets based on various ML setups?
ML has application in a range of topical areas. In the healthcare space, ML is applied to disease diagnostics, from various multimodal data as inputs; medicine creation and research; genetic analysis, and other ways to improve healthcare. It enables the efficient understanding of information that may not have been acquire-able before. From AI’s earliest days, it has been used in the medical industry.
Ashutosh Srivastava, Aman Anand, Praveen Kumar, and Sudeshna Chakraborty’s “Comprehensive Review on Applications of Machine Learning in Healthcare” (Ch. 2) opens with a definition of machine learning as…
Machine learning is trained from “previously applied computations to produce reliability, repeatable decisions and results” (Srivastava, Anand, Kumar, & Chakraborty, 2022, p. 25). Such findings can inform healthcare and related decision-making, disease detection and diagnoses, forecasting, health interventions, health service planning and delivery, budgeting, policymaking, and other health-related decisions. To achieve such ends in healthcare, the analysts need access to private and protected health data, de-identified health data, summary health data, or historical population health or research data, and they need channels through which they can share the data to inform human health policies and decisions. In many cases, such health data, which is high-dimensional, may be analyzed in connection with demographic and other data as well. These new technological approaches are predicted to add “an additional 15.7 trillion to the world economy by 2030, and the largest effect will be on health care” (Srivastava, Anand, Kumar, & Chakraborty, 2022, p. 27).
This team summarizes some known applications of ML in this space, such as how to improve public health clinics in rural areas, how to improve radiology imaging interpretations, drug discovery, drug experiments, biomedicine, and applying natural language processing (NLP) to patient feedback and clinical notes.
They describe the use of a Multi-Armed Bandit (MAB) approach in which the AI works to find the optimal settings for the highest reward (based on the computational objective or objectives). They highlight Generative Adversarial Networks (GANs), which both generate models or artifacts and then analyze those against standards for more effective creation. There is Bayesian Optimization, Graph Cuts, and Markov Random Fields (Srivastava, Anand, Kumar, & Chakraborty, 2022). The authors describe the various approaches in fairly straightforward ways although some elaboration may benefit some examples.
A widely known application is segmentation of medical images, such as between healthy and unhealthy tissue. Computer-aided Detection (CADe) and Computer-aided Diagnostic (CADx) techniques are widely used in medicine. There are dedicated tools for particular detection of particular health phenomenon. (Srivastava, Anand, Kumar, & Chakraborty, 2022) This work is an engaging one, but it is not clear how close it is to describing the current state of the art of ML in healthcare given so many advances and the lag with academic research, writing, and publishing.
Mahua Nandy Pal and Minakshi Banerjee’s “Retinal Vessel Enhancement Evaluation on DRIVE Dataset from CNN Performance” (Ch. 3) describes the segmentation of deep vessels in images of the human retina based on image enhancement techniques and assesses which computational approach is most effective based on metrics. The health of the retina, which may be visibly observed, may show perhaps systemic health issues like hypertension (high blood pressure) and diabetes (high blood sugar), or more localized ocular health issues. This work evaluates Contrast Limited Adaptive Histogram Equalization (CLAHE), Adaptive Gamma Correction (AGC), Tophat transformation, and various sequential combinations of analyses to identify the optimal performance, based on the Digital Retinal Image for Vessel Extraction (DRIVE) dataset (by the Image Sciences Institute).
After a light description of the various technological methods, the researchers assess the approaches based on common evaluative metrics like true positives, false positives, true negatives, and false negatives in terms of vessel classification. There are also other variables captured such as “sensitivity, specificity, accuracy and AUC” (area under the curve) as important evaluative metrics as well (Pal & Banerjee, 2022, pp. 51-52), typical variables for predictive classifiers. The research is described precisely and methodically. As in many ML competitions, the competitive advantages may be seen in very small margins, such as a percent of a percent; those margins matter when handling larger datasets and perhaps for bragging rights. The researchers found that using convolutional neural network (CNN) offered better classification metrics “to obtain from better-enhanced input image datasets” (Pal & Banerjee, 2022, p. 60). The researchers also found that “among all different possibilities, consecutive application of Tophat transformation, AGC and CLAHE is the most proficient with respect to vessel enhancement while applied to retinal fundus images” (Pal & Banerjee, 2022, p. 60).
Arjun Singh and Punit Gupta’s “Machine Learning for Network Analysis” (Ch. 4) examines various approaches for machine learning networking (MLN) for more secure and efficient distributed computing. Machine learning…
A typical sequence in ML setup for organizational studies include the following: “issue detailing, information assortment, information investigation, model development, model approval, arrangement, and surmising” (Singh & Gupta, 2022, p. 70). The preparation cycle is known to be fairly onerous. Pre-processing and data cleaning may involve data standardization, discretizing data, and dealing with missing data (p. 72). Annotations should also be removed. Mistakes in the setup and running can lead to an “inadmissible learning model” and wrong modeling and erroneous findings (Singh & Gupta, 2022, p. 71). Results are tested to see if the model made the most of the data, not overfitting, not underfitting.
This work explores the application of ML to assess the performance of various electronic networks: wired, wireless; ad hoc; mobile, sensor, IoT, and others. Assessing networks involves looking at various issues like “routing, clustering, power efficiency, life spam (sic), fault detection, security, (and) intruder detection” among others (Singh & Gupta, 2022, pp. 73-74). Some applications focus on particular aspects of particular networks, such as how to identify nodes of interest in various networks; how to recognize location-based activities in smart homes from network activity; forest fire detection from monitoring wireless sensor networks; event detection during disaster management; and other practical applications. There are various named algorithms for identifying clusters in networks in collision-free ways (so that a node is not identified twice). In real-world conditions, the algorithms all have their strengths and weaknesses. The researchers describe the ML process for training their new network model: “1) selection of features and marking of outputs, 2) collection of samples, 3) offline training and 4) online classification” (Singh & Gupta, 2022, p. 78).
At the time of this review, climate change and its more extreme weather have resulted in some mass-scale disasters, with major flooding in Pakistan across two million acres of that country. Other locales around the world are flooded, too, with some suffering catastrophic damage to their potable water systems.
S.M. Kodirov and Pooja’s “Machine Learning in Hydrology” (Ch. 5) focuses on the application of ML to world hydrological practice, including the analysis of hydrometeorological information. This work focuses on hydrology and meteorology, and related tools, equipment, and methods, in this sector for a general informed audience. This work seems to use a strong exposition of the present field in order to encourage advances into the future. The researchers specify that all models are incomplete representations of reality. So too with hydrological modeling, which generally is used to address “design of sewerage systems; design of drainage systems; design of small dams and reservoirs” (Kodirov & Pooja, 2022, p. 85).
The researchers introduce readers to common terminology. They summarize Manning’s formula, the unit hydrograph method of the 1930s, more complex dynamic modeling of systems in the 1950s, sporadic flooding and precipitation-runoff modeling in the 1960s, stochastic models in the 1970s, fully distributed physical models in the 1980s, large-scale hydrological models in the 1990s (Kodirov & Pooja, 2022, pp. 86-87). The researchers differentiate between physical (theoretical) and hydrodynamic models, conceptual models (models that combine theory and research), and empirical ones (based on research and observables) (Kodirov & Pooja, 2022, p. 88).
“Lumped models” refer to entire river basins as one unit in the model; “fully distributed models” refer to “physically based models (the entire domain is distributed completely over elementary areas by using grids),” “semi-distributed models” focus on “sub-basins or sub-catchments and hydrotops” for the simulation, and then there are regional divides, such as between global and regional models based on scale (planetary and smaller areas) (Kodirov & Pooja, 2022, pp. 89 - 90).
Different models should be applied to acquire different applied information. They write:
How the models are harnessed is important, too. It is important to ensure that a model is sensitive for unknown parameters. The model has to be calibrated with a sensitivity analysis to ensure that model parameters do not lead to overfitting or underfitting to the local data. Parameter changes to the model may have excessive or undesirable effects on the simulated variable.
The hydrological cycles on Earth generally include the following elements: “precipitation, snow melting, physical evaporation, biological evaporation (through plants), water retention by forest canopy, surface runoff, subsoil runoff, base flow, ground runoff, river runoff, melting glaciers, (and) biochemical processes in soil and water” (Kodirov & Pooja, 2022, p. 92). Further: “The two most important processes that affect runoff formation are precipitation and evaporation (physical and biological), also known as evapotranspiration” (Kodirov & Pooja, 2022, p. 92).
In the field, there are hydrological databases of river flow data from various time periods and locations. The hydrological information may be explored for different historical patterns and insights.
This work seems to stop in the middle, without actual analysis of machine learning applications in the hydrology space. Perhaps the idea was to offer a sense of the field and to invite practitioners to do more of the thinking and planning work to achieve some new ML applications.
Ruchi Yadav, Ashish Gupta, Rashmi Priyadarshini, Parma Nand, and Agha Asim Husain’s “Machine Learning Applications in Smart Sensors” (Ch. 6) focuses on devices that capture information from the larger environment, especially wearables. Smart sensors are of two general categories, those that sense physical and chemical information, and those that capture visual (p. 97). Wearable smart sensors measure various health parameters to inform on health. This work summarizes several case studies: one is about sensors in shoes to analyze gait for diabetic foot monitoring. The GAIT system “consists of accelerometers, gyroscopes, inclinometers that uses (sic) a plate of multi camera-based system, magneto-resistive type of sensors, vertical ground reaction force (vGRF) that is the force between the foot and ground” (Yadav, Gupta, Priyadarshini, Nand, & Husain, 2022, p. 100). The pressure sensor arrays may inform how insoles are designed. Indeed, there are practical and cost considerations in these discussions. Another case explores blood pressure estimation from “photo plethysmo gram” signal. Another case involves wearables that can detect real-time heart attacks (myocardial infarctions). There are real-time “smart-digital stethoscope” systems that provide readings that may detect cardiovascular diseases. The research involves measuring data accuracy based on various competing ML algorithms.
Rashi Srivastava, Priya Tyagi, Parma Nand, Ankur Sharma, Niraj Kumar Jha, and Saurabh Kumar Jha’s “The Role of Machine Learning in the Food Industry” (Ch. 7) begins with the observation of the high “variability in raw materials and ingredients, stochastic yield and demand” and other factors that influence the actual yields as compared anticipated ones (p. 121). Awareness of the variability may help all those working in food production make potential adjustments and pivots that may result in higher yields or to adapt to difficult impending circumstances. The researchers write:
Gouri Sankar Mishra, Parma Nand, Rani Astya, Nitin Rakesh, Pradeep Kumar Mishra, and Sasmita Mishra’s “Phrase Based Machine Translation of English Text to Indian Sign Language Using Order Scoring Algorithm” (Ch. 8) introduces a method used to improve the translating of English text phrases into Indian sign language, for those who are deaf or hard-of-hearing. At heart is an order scoring algorithm to lessen the error-proneness of the machine translations. The researchers tested their system with announcements on the Indian Railway Platform and general conversations, in this preliminary research project.
Another space in which ML is seen to have an important effect is in stem cell and tissue engineering. These endeavors are relevant to human (and animal) health because of the capability to treat various diseases through the regeneration of damaged tissue or cells, through bio-engineering.
In their abstract, Rashi Srivastava, Parma Nand, Dhruv Kumar, Ankur Sharma, Niraj Kumar Jha, and Saurabh Kumar Jha’s “Stem Cell and Tissue Engineering: Application of Machine Learning” (Ch. 9) write that ML helps “accelerate the development of novel materials, processes and techniques”; it plays important roles in “analysis of cell-to-cell network, generation of blueprint for tissues, reprogramming the stem cells, deciphering and analyzing the interaction and effect of individual donor characteristics and their growth dynamics, for maturation and fabrication of 3D bio-printed tissues and organs, and others” (p. 152).
Advanced data analysis has come to the fore in this space because of advances in the field:
Stem cell and tissue engineering’s focus is “to foster a desired cell/tissue for a specific target individual with the help of pheno-genotypic information of the target cells” (p. 154). The new approaches are important “in drug screening, disease modeling, regenerative medicine, and precision medicine” (p. 155). For the ML systems to work, some 10,000 images are required for an accurate learning model. A diagram shows how induced pluripotent stem cells are extracted from a person, placed in a transmitted light microscope, benchmarked and calibrated, rendered into quantitative absorbance images, analyzed with “deep/machine learning algorithms,” and then analyzed to “predict identity, function and outlier status” (p. 156). Such assessments may be used to predict how a human heart may respond to the application of particular drugs, for example (based on cellular analysis). These technologies may be used for “organ modeling” or biomechanical modeling, for health interventions. Tissues may be engineered ex vivo (or outside the human body). Indeed, there are high hopes for the harnessing of ML in the bioengineering space for human betterment.
3D point clouds may be used to create 3D designs in space and other applications. Souvik Das, Asheesh Girdharwal, J. Maiti, and O.B. Krishna’s “Reconstruction of 3D Point Cloud--Based on the Sequence of Images” (Ch. 10) describes a computational approach in which a sequence of images are transcoded into 3D point clouds. The sequential pictures, such as videos of a structure, require slow motion, precise visual captures, and a “complete view” of the space of interest (p. 165). Various features are extracted to depict a 3D structure. This application may be applied to “industrial automation, Virtual Reality, CAD design” (p. 165), VR trainings in high-risk industries, and other applications. To achieve this end, there are some “major algorithms like Structure from Motion (SfM), Multiview stereo (MVS)” and Poisson surface reconstruction (p. 168).
The general steps to the tested algorithm involve visual extraction of video frames from the video, matching features across the frames to “find out corner points in the images,” then matching features in the consecutive images (Das, Girdharwal, Maiti, & Krishna, 2022, p. 169). A subsequent step verifies the geometries.
The respective observed structures are reconstructed in three dimensions. The researchers write that the reconstruction starts with an initial pair of images, and new pictures are “added or registered to the 3-d reconstructed scene” (Das, Girdharwal, Maiti, & Krishna, 2022, p. 173). Feature matching enables the triangulation of points in the already registered images. The program takes as input a set of images and “returns sparse point cloud” (p. 174), which can be transcoded to a CAD-importable model, which increases designer or engineer or artist or architect efficiency. Some “very fine details” are missing in the process (p. 174).
Hemendra Gupta and Anviti Gupta’s “Application of Machine Learning in Financial Services” (Ch. 11) lists a variety of ways that ML enhances the financial services industry. ML is important “in the banking ecosystem, such as loan approvals, assessing credit scores, managing assets, identifying early bad debts and underwriting risks” (p. 183). Social robots help provide customer service. Sentiment analysis is applied in “portfolio and hedge management services” (p. 183). Call centers are automated. Paperwork is automated for certain processes. Social robots provide wealth advisory services based on the particular contexts of the clients. The network is better protected against cybercrime adversaries, and there are stronger fraud protections in place. Algorithms are used for trading. AI creates legal and written contents. There are opportunities for custom solutions as well (p. 191).
This research study collected two types of data. One involved various psychometric measures for “attention, mental agility and problem solving ability at intervals of 1 hour till the participant is 24 hours sleep deprived” (Das, Pratyush, Das, Maiti, & Krishna, 2022, p. 199). The second type of data collected involved the eye-tracking data recorded from a device worn during the experiment.
The collected data was then analyzed using a Hidden Markov model:
The research team found that there was “a decrease in all 3 parameters with an increase in sleep deprivation,” so drops in attention, mental agility, and problem-solving. Problem-solving was the “most heavily affected by sleep deprivation” (Das, Pratyush, Das, Maiti, & Krishna, 2022, p. 205). This work shows well that solid research designs are important, so the machine learning analytics of the data have relevance and impact.
Figure 5. Artificial Neural Network Illustration
Shalin Hai-Jew works as an instructional designer / researcher at Kansas State University. Her email is shalin@ksu.edu.
Machine Learning Algorithms for Engineering Applications: Future Trends and Research Directions
Prasenjit Chatterjee, Parmanand Astya, Sudeshna Chakraborty, and Pooja, Editors
Nova Science Publishing
2022
208 pp.
Prasenjit Chatterjee, Parmanand Astya, Sudeshna Chakraborty, and Pooja, Editors of Machine Learning Algorithms for Engineering Applications: Future Trends and Research Directions (2022), offer an edited text that bridges the reader to practical applications of machine learning. As computer scientists make advancements in their field, engineers and area-experts are critical in harnessing those advances for real-world solutions.
Preface
Machine learning (ML) is a sub-element of artificial intelligence (AI), informally defined as synthetic intelligence inspired by biological human / animal / plant sources. ML expands what is knowable given sufficient and relevant information, from which data models are made; statistical and mathematical analyses are applied to understand potentially relevant data patterns.
Readers of the text are advised to know “R, CUDA, Mahout, Pytorch, TensorFlow, and CoLab” to work the examples (Chatterjee, Astya, Chakraborty, & Pooja, 2022, p. x). While programming is not apparently required, that knowledge cannot hurt. ML tools may run on local machines or remote ones. Machine learning platforms often come with “extensive libraries, a user-friendly GUI, and support for popular programming languages” (Chatterjee, Astya, Chakraborty, & Pooja, 2022, pp. x-xi).
There are three general paradigms of machine learning: (1) supervised machine learning (with labeled data), (2) unsupervised machine learning (without labeled data), and (3) reinforcement learning (a method that does not need labeled data, which applies dynamic programming applying an inexact Markov decision process to both explore and exploit data to optimize targeted learning). More on the approaches and related algorithms follow later on. This is done without dedicated or specific programming, but there is logic in how to set up ML for particular datasets and explorations. Others in their ML typography include ensemble learning (combining multiple AI approaches) and various types of artificial neural networks.
In a general sense, machine learning has been applied to predictive analytics, such as classifications. It has been applied to object recognition in digital images. It is applied for speech-to-text translation. ML enables the generation of natural language expressions of various types (such as humor, writing in various genres, and others).
Machine Learning Tools
Parveen Mor and Sonia Chhabra’s “Tools for Machine Learning” (Ch. 1) describe machine learning as a “technique,” with ways for computers to find data patterns and “learn” while “not being expressly programmed” (p. 1). This team offers advice to those getting started with machine learning:
Before developing machine learning applications, it’s important to pick a machine learning tool that has intensive libraries, a nice interface, and support for common programming languages. (Mor & Chhabra, 2022, p. 2)
Common machine learning languages include Python, C++, and R, among others. Machine learning applications include NumPy, Scikit-learn, NLTK (Natural Language Toolkit in Python), and TensorFlow, likely all recognizable names.
Supervised machine learning takes as input labeled data, so the intelligent agents examine information based on historical knowledge. Unsupervised machine learning takes as inputs unlabeled data, but these approaches enable various clustering based on observed associations. Reinforcement machine learning also works without labeled data, and the methods in this paradigm work flexibly to maximize targeted learning towards a “cumulative reward” of outcomes. The reinforcement approaches enable learning from the available data to optimize desired outcomes, using a Markov decision process (without defining an exact mathematical model, often given the large sizes of Markov decision processes). They write of ML that it requires “deep neural networks and machine learning methods to assist programming” (Mor & Chhabra, 2022, p. 9).
The researchers keep the writing at a general level and offer comparative strengths and weaknesses of the respective technologies. They note that in earlier times, there were command-line ways to interact with machine learning, but many of the available platforms are moving from low-code to no-code requirements. They mention KNIME Analytics Platform as a free and open source downloadable tool. They point to TensorFlow as a free and open source one, too. There are cloud services like Google Colab (Python-supported cloud service) that enables ML analytics. They also mention Apache Mahout, Accord.Net network, Shogun, Keras.io, Apache Singa, Apache Spark MLlib, and others. There are serious efforts to democratize access to machine learning.
This work focuses on free and low-cost solutions. They do mention RapidMiner, but they also show the relatively expensive cost for non-educational uses of the tool. They summarize their descriptions of the respective technologies in a data table, with costs, the computer languages integrated, and the available algorithms at the time of the publication. The table is not comprehensive, and various importable “packages” and such were not necessarily addressed. One capability was labeled as “dimensionality decreases” when what might have been meant was “dimensionality reduction.” Another capability was labeled as “data preparation,” but that term is very general and could mean any number of things.
It is encouraging to see quite a number of gateways to bridge to machine learning. There are numerous opportunities for experimentation, with public datasets, with labeled and unlabeled data.
What would enhance this chapter would be some more real-world examples of ML analytics. It would help to have some examples of how to approach a dataset and set it up for ML to explore particular dimensions of that data. In other words: How exploitable are large datasets based on various ML setups?
ML in Healthcare
ML has application in a range of topical areas. In the healthcare space, ML is applied to disease diagnostics, from various multimodal data as inputs; medicine creation and research; genetic analysis, and other ways to improve healthcare. It enables the efficient understanding of information that may not have been acquire-able before. From AI’s earliest days, it has been used in the medical industry.
Ashutosh Srivastava, Aman Anand, Praveen Kumar, and Sudeshna Chakraborty’s “Comprehensive Review on Applications of Machine Learning in Healthcare” (Ch. 2) opens with a definition of machine learning as…
… the use of artificial intelligence which gives the system the expertise to master and upgrade themselves automatically from experience in the absence of being clearly programmed. (Srivastava, Anand, Kumar, & Chakraborty, 2022, p. 25)
Machine learning is trained from “previously applied computations to produce reliability, repeatable decisions and results” (Srivastava, Anand, Kumar, & Chakraborty, 2022, p. 25). Such findings can inform healthcare and related decision-making, disease detection and diagnoses, forecasting, health interventions, health service planning and delivery, budgeting, policymaking, and other health-related decisions. To achieve such ends in healthcare, the analysts need access to private and protected health data, de-identified health data, summary health data, or historical population health or research data, and they need channels through which they can share the data to inform human health policies and decisions. In many cases, such health data, which is high-dimensional, may be analyzed in connection with demographic and other data as well. These new technological approaches are predicted to add “an additional 15.7 trillion to the world economy by 2030, and the largest effect will be on health care” (Srivastava, Anand, Kumar, & Chakraborty, 2022, p. 27).
This team summarizes some known applications of ML in this space, such as how to improve public health clinics in rural areas, how to improve radiology imaging interpretations, drug discovery, drug experiments, biomedicine, and applying natural language processing (NLP) to patient feedback and clinical notes.
They describe the use of a Multi-Armed Bandit (MAB) approach in which the AI works to find the optimal settings for the highest reward (based on the computational objective or objectives). They highlight Generative Adversarial Networks (GANs), which both generate models or artifacts and then analyze those against standards for more effective creation. There is Bayesian Optimization, Graph Cuts, and Markov Random Fields (Srivastava, Anand, Kumar, & Chakraborty, 2022). The authors describe the various approaches in fairly straightforward ways although some elaboration may benefit some examples.
A widely known application is segmentation of medical images, such as between healthy and unhealthy tissue. Computer-aided Detection (CADe) and Computer-aided Diagnostic (CADx) techniques are widely used in medicine. There are dedicated tools for particular detection of particular health phenomenon. (Srivastava, Anand, Kumar, & Chakraborty, 2022) This work is an engaging one, but it is not clear how close it is to describing the current state of the art of ML in healthcare given so many advances and the lag with academic research, writing, and publishing.
Using AI to Better View Retinal Vessels
Mahua Nandy Pal and Minakshi Banerjee’s “Retinal Vessel Enhancement Evaluation on DRIVE Dataset from CNN Performance” (Ch. 3) describes the segmentation of deep vessels in images of the human retina based on image enhancement techniques and assesses which computational approach is most effective based on metrics. The health of the retina, which may be visibly observed, may show perhaps systemic health issues like hypertension (high blood pressure) and diabetes (high blood sugar), or more localized ocular health issues. This work evaluates Contrast Limited Adaptive Histogram Equalization (CLAHE), Adaptive Gamma Correction (AGC), Tophat transformation, and various sequential combinations of analyses to identify the optimal performance, based on the Digital Retinal Image for Vessel Extraction (DRIVE) dataset (by the Image Sciences Institute).
After a light description of the various technological methods, the researchers assess the approaches based on common evaluative metrics like true positives, false positives, true negatives, and false negatives in terms of vessel classification. There are also other variables captured such as “sensitivity, specificity, accuracy and AUC” (area under the curve) as important evaluative metrics as well (Pal & Banerjee, 2022, pp. 51-52), typical variables for predictive classifiers. The research is described precisely and methodically. As in many ML competitions, the competitive advantages may be seen in very small margins, such as a percent of a percent; those margins matter when handling larger datasets and perhaps for bragging rights. The researchers found that using convolutional neural network (CNN) offered better classification metrics “to obtain from better-enhanced input image datasets” (Pal & Banerjee, 2022, p. 60). The researchers also found that “among all different possibilities, consecutive application of Tophat transformation, AGC and CLAHE is the most proficient with respect to vessel enhancement while applied to retinal fundus images” (Pal & Banerjee, 2022, p. 60).
ML and Network Analysis
Arjun Singh and Punit Gupta’s “Machine Learning for Network Analysis” (Ch. 4) examines various approaches for machine learning networking (MLN) for more secure and efficient distributed computing. Machine learning…
… tries to construct algorithms and models that can learn to make judgments directly from data without following predefined rules. Present machine learning algorithms generally fall into three categories: supervised learning (SL), unsupervised learning (USL), and reinforcement learning (RL). All the more explicitly, SL algorithms figure out how to lead order or relapse undertakings from named information, while USL algorithms have practical experience in arranging the example sets into various gatherings (i.e., clusters) with unlabelled (sic) information. In RL algorithms, specialists figure out how to locate the best activity arrangement to boost the cumulated reward (i.e., target work) by cooperating with the climate. The most recent discoveries, including deep learning (DL) move to learn, and generative adversarial networks (GAN), likewise give possible examination and application bearings in an unfathomable design. (Singh & Gupta, 2022, p. 64)
There is a human role in this at every stage. There is the setup of the ML approaches, the cleaning of the data, the analyses of the outputs, all of which require expertise and analytical capabilities. The state of the data influences what the ML outputs will be, and such content will not reoccur exactly in the same way (frequencies, types) in the future. They write: “Consequently, every student must epitomize some information or presumptions past the information it’s provided so as, to sum up past it. ML can be applied in administered, semi-regulated, and solo modes” (Singh & Gupta, 2022, p. 66). The models have to be fine-tuned for more accurate learning and to more effectively forecast into the future, which has not yet arrived. As to what ML will be used, there are considerations of computational costs and of available insights.
Figure 2. Wired and Wireless and Mixed Networks
This work explores the application of ML to assess the performance of various electronic networks: wired, wireless; ad hoc; mobile, sensor, IoT, and others. Assessing networks involves looking at various issues like “routing, clustering, power efficiency, life spam (sic), fault detection, security, (and) intruder detection” among others (Singh & Gupta, 2022, pp. 73-74). Some applications focus on particular aspects of particular networks, such as how to identify nodes of interest in various networks; how to recognize location-based activities in smart homes from network activity; forest fire detection from monitoring wireless sensor networks; event detection during disaster management; and other practical applications. There are various named algorithms for identifying clusters in networks in collision-free ways (so that a node is not identified twice). In real-world conditions, the algorithms all have their strengths and weaknesses. The researchers describe the ML process for training their new network model: “1) selection of features and marking of outputs, 2) collection of samples, 3) offline training and 4) online classification” (Singh & Gupta, 2022, p. 78).
ML and Hydrology
At the time of this review, climate change and its more extreme weather have resulted in some mass-scale disasters, with major flooding in Pakistan across two million acres of that country. Other locales around the world are flooded, too, with some suffering catastrophic damage to their potable water systems.
S.M. Kodirov and Pooja’s “Machine Learning in Hydrology” (Ch. 5) focuses on the application of ML to world hydrological practice, including the analysis of hydrometeorological information. This work focuses on hydrology and meteorology, and related tools, equipment, and methods, in this sector for a general informed audience. This work seems to use a strong exposition of the present field in order to encourage advances into the future. The researchers specify that all models are incomplete representations of reality. So too with hydrological modeling, which generally is used to address “design of sewerage systems; design of drainage systems; design of small dams and reservoirs” (Kodirov & Pooja, 2022, p. 85).
The researchers introduce readers to common terminology. They summarize Manning’s formula, the unit hydrograph method of the 1930s, more complex dynamic modeling of systems in the 1950s, sporadic flooding and precipitation-runoff modeling in the 1960s, stochastic models in the 1970s, fully distributed physical models in the 1980s, large-scale hydrological models in the 1990s (Kodirov & Pooja, 2022, pp. 86-87). The researchers differentiate between physical (theoretical) and hydrodynamic models, conceptual models (models that combine theory and research), and empirical ones (based on research and observables) (Kodirov & Pooja, 2022, p. 88).
“Lumped models” refer to entire river basins as one unit in the model; “fully distributed models” refer to “physically based models (the entire domain is distributed completely over elementary areas by using grids),” “semi-distributed models” focus on “sub-basins or sub-catchments and hydrotops” for the simulation, and then there are regional divides, such as between global and regional models based on scale (planetary and smaller areas) (Kodirov & Pooja, 2022, pp. 89 - 90).
Different models should be applied to acquire different applied information. They write:
For example, if we are interested in runoff from a given particular rain that is predicted to occur in two to three days, apparently we should use an unallocated model that will quickly but not very accurately estimate the total runoff. On the other hand, if we are interested not only in river flow but also in the amount of agricultural production or in assessing how land use change will affect water discharge, it makes sense to apply models such as SWIM and SWAT. (Kodirov & Pooja, 2022, p. 91)
How the models are harnessed is important, too. It is important to ensure that a model is sensitive for unknown parameters. The model has to be calibrated with a sensitivity analysis to ensure that model parameters do not lead to overfitting or underfitting to the local data. Parameter changes to the model may have excessive or undesirable effects on the simulated variable.
The hydrological cycles on Earth generally include the following elements: “precipitation, snow melting, physical evaporation, biological evaporation (through plants), water retention by forest canopy, surface runoff, subsoil runoff, base flow, ground runoff, river runoff, melting glaciers, (and) biochemical processes in soil and water” (Kodirov & Pooja, 2022, p. 92). Further: “The two most important processes that affect runoff formation are precipitation and evaporation (physical and biological), also known as evapotranspiration” (Kodirov & Pooja, 2022, p. 92).
In the field, there are hydrological databases of river flow data from various time periods and locations. The hydrological information may be explored for different historical patterns and insights.
This work seems to stop in the middle, without actual analysis of machine learning applications in the hydrology space. Perhaps the idea was to offer a sense of the field and to invite practitioners to do more of the thinking and planning work to achieve some new ML applications.
ML and Wearable Smart Sensors
Ruchi Yadav, Ashish Gupta, Rashmi Priyadarshini, Parma Nand, and Agha Asim Husain’s “Machine Learning Applications in Smart Sensors” (Ch. 6) focuses on devices that capture information from the larger environment, especially wearables. Smart sensors are of two general categories, those that sense physical and chemical information, and those that capture visual (p. 97). Wearable smart sensors measure various health parameters to inform on health. This work summarizes several case studies: one is about sensors in shoes to analyze gait for diabetic foot monitoring. The GAIT system “consists of accelerometers, gyroscopes, inclinometers that uses (sic) a plate of multi camera-based system, magneto-resistive type of sensors, vertical ground reaction force (vGRF) that is the force between the foot and ground” (Yadav, Gupta, Priyadarshini, Nand, & Husain, 2022, p. 100). The pressure sensor arrays may inform how insoles are designed. Indeed, there are practical and cost considerations in these discussions. Another case explores blood pressure estimation from “photo plethysmo gram” signal. Another case involves wearables that can detect real-time heart attacks (myocardial infarctions). There are real-time “smart-digital stethoscope” systems that provide readings that may detect cardiovascular diseases. The research involves measuring data accuracy based on various competing ML algorithms.
ML and the Food Industry
Rashi Srivastava, Priya Tyagi, Parma Nand, Ankur Sharma, Niraj Kumar Jha, and Saurabh Kumar Jha’s “The Role of Machine Learning in the Food Industry” (Ch. 7) begins with the observation of the high “variability in raw materials and ingredients, stochastic yield and demand” and other factors that influence the actual yields as compared anticipated ones (p. 121). Awareness of the variability may help all those working in food production make potential adjustments and pivots that may result in higher yields or to adapt to difficult impending circumstances. The researchers write:
Machine learning is already presented as an asset for the food industry in a number of the researches like, predicting and managing the sales, forecasting the yield, and others. Furthermore, Machine Learning algorithms can prove themselves: to conclude interpretable rules for classifying samples regardless of the non-linearity; to extract functioning human knowledge from a list of set examples, and help to discover the degree of influence of each objective attribute of the assessed food. (Srivastava, Tyagi, Nand, Sharma, Jha, & Jha, 2022, p. 122)
Some common uses of machine learning involve assessing for food quality, such as assessments that evaluate geographical origins of the product, botanical origins, pesticide presence, authentication based on trace elements, sensory quality evaluation, and shelf life, among others. ML is also used in the food supply chain management, the management of food waste, and restaurant management. This work is informed by a review of the literature. As such, these examples may benefit others’ applications of ML in their respective sectors.
Figure 3. An AI's Idea of Food
Machine Translation from English Text to Indian Sign Language
Gouri Sankar Mishra, Parma Nand, Rani Astya, Nitin Rakesh, Pradeep Kumar Mishra, and Sasmita Mishra’s “Phrase Based Machine Translation of English Text to Indian Sign Language Using Order Scoring Algorithm” (Ch. 8) introduces a method used to improve the translating of English text phrases into Indian sign language, for those who are deaf or hard-of-hearing. At heart is an order scoring algorithm to lessen the error-proneness of the machine translations. The researchers tested their system with announcements on the Indian Railway Platform and general conversations, in this preliminary research project.
ML and Stem Cell and Tissue Bio-engineering
Another space in which ML is seen to have an important effect is in stem cell and tissue engineering. These endeavors are relevant to human (and animal) health because of the capability to treat various diseases through the regeneration of damaged tissue or cells, through bio-engineering.
In their abstract, Rashi Srivastava, Parma Nand, Dhruv Kumar, Ankur Sharma, Niraj Kumar Jha, and Saurabh Kumar Jha’s “Stem Cell and Tissue Engineering: Application of Machine Learning” (Ch. 9) write that ML helps “accelerate the development of novel materials, processes and techniques”; it plays important roles in “analysis of cell-to-cell network, generation of blueprint for tissues, reprogramming the stem cells, deciphering and analyzing the interaction and effect of individual donor characteristics and their growth dynamics, for maturation and fabrication of 3D bio-printed tissues and organs, and others” (p. 152).
Advanced data analysis has come to the fore in this space because of advances in the field:
The advent of induced pluripotent stem cells technologies, organoid technologies and CRISPR/Cas9 genetic engineering; (sic) the stated field is now significantly expanded itself to environ biomimetic engineering. The study and analysis of genomic, proteomic, epigenomic, metabolomic, transcriptomic and others is becoming an essential part to discover the efficiency, specificity, safety of the engineered cells and tissues with respect to the patient. (Srivastava, Nand, Kumar, Sharma, Jha, & Jha, 2022, p. 154)
Stem cell and tissue engineering’s focus is “to foster a desired cell/tissue for a specific target individual with the help of pheno-genotypic information of the target cells” (p. 154). The new approaches are important “in drug screening, disease modeling, regenerative medicine, and precision medicine” (p. 155). For the ML systems to work, some 10,000 images are required for an accurate learning model. A diagram shows how induced pluripotent stem cells are extracted from a person, placed in a transmitted light microscope, benchmarked and calibrated, rendered into quantitative absorbance images, analyzed with “deep/machine learning algorithms,” and then analyzed to “predict identity, function and outlier status” (p. 156). Such assessments may be used to predict how a human heart may respond to the application of particular drugs, for example (based on cellular analysis). These technologies may be used for “organ modeling” or biomechanical modeling, for health interventions. Tissues may be engineered ex vivo (or outside the human body). Indeed, there are high hopes for the harnessing of ML in the bioengineering space for human betterment.
Reconstructing 3D Point Clouds with Machine Learning
3D point clouds may be used to create 3D designs in space and other applications. Souvik Das, Asheesh Girdharwal, J. Maiti, and O.B. Krishna’s “Reconstruction of 3D Point Cloud--Based on the Sequence of Images” (Ch. 10) describes a computational approach in which a sequence of images are transcoded into 3D point clouds. The sequential pictures, such as videos of a structure, require slow motion, precise visual captures, and a “complete view” of the space of interest (p. 165). Various features are extracted to depict a 3D structure. This application may be applied to “industrial automation, Virtual Reality, CAD design” (p. 165), VR trainings in high-risk industries, and other applications. To achieve this end, there are some “major algorithms like Structure from Motion (SfM), Multiview stereo (MVS)” and Poisson surface reconstruction (p. 168).
The general steps to the tested algorithm involve visual extraction of video frames from the video, matching features across the frames to “find out corner points in the images,” then matching features in the consecutive images (Das, Girdharwal, Maiti, & Krishna, 2022, p. 169). A subsequent step verifies the geometries.
The respective observed structures are reconstructed in three dimensions. The researchers write that the reconstruction starts with an initial pair of images, and new pictures are “added or registered to the 3-d reconstructed scene” (Das, Girdharwal, Maiti, & Krishna, 2022, p. 173). Feature matching enables the triangulation of points in the already registered images. The program takes as input a set of images and “returns sparse point cloud” (p. 174), which can be transcoded to a CAD-importable model, which increases designer or engineer or artist or architect efficiency. Some “very fine details” are missing in the process (p. 174).
ML and Financial Services
Hemendra Gupta and Anviti Gupta’s “Application of Machine Learning in Financial Services” (Ch. 11) lists a variety of ways that ML enhances the financial services industry. ML is important “in the banking ecosystem, such as loan approvals, assessing credit scores, managing assets, identifying early bad debts and underwriting risks” (p. 183). Social robots help provide customer service. Sentiment analysis is applied in “portfolio and hedge management services” (p. 183). Call centers are automated. Paperwork is automated for certain processes. Social robots provide wealth advisory services based on the particular contexts of the clients. The network is better protected against cybercrime adversaries, and there are stronger fraud protections in place. Algorithms are used for trading. AI creates legal and written contents. There are opportunities for custom solutions as well (p. 191).
Understanding Effects of Sleep Deprivation Detection via ML
Figure 4. "Sleep Deprivation, Fairytale" Prompt on CrAIyon AI for Art Drawing
Souvik Das, Parimal Pratyush, Debjoy Das, J. Maiti, and O.B. Krishna’s “Eye-Tracking Data as a Way to Detect Sleep Deprivation in an Individual, Based on Attention, Mental Agility, and Problem-Solving” (Ch. 12) explores potential relationships between sleepiness and particular aspects of human mental or cognitive functioning. This work uses eye-tracking data as an input, including visuals of “pupil dilation, microsaccades (magnitude and speed) and fixation” (p. 195). It is thought that the ability to use eye-tracking to detect employee fatigue in industrial operations, transport services, healthcare, and other fields that worker safety and performance may be enhanced. Sleep deprivation is generally measured based on the person’s “performance, mood and objective sleepiness” (p. 197).
This research study collected two types of data. One involved various psychometric measures for “attention, mental agility and problem solving ability at intervals of 1 hour till the participant is 24 hours sleep deprived” (Das, Pratyush, Das, Maiti, & Krishna, 2022, p. 199). The second type of data collected involved the eye-tracking data recorded from a device worn during the experiment.
The collected data was then analyzed using a Hidden Markov model:
The model moves from the current state to the next state, which may be the same state, according to some probability distribution that depends only on the current state…After every transition, the model gives an observation whose distribution depends only on the current state…The states that generate these observations are called hidden states as they are unknown to the user. (Das, Pratyush, Das, Maiti, & Krishna, 2022, p. 204)
The research team found that there was “a decrease in all 3 parameters with an increase in sleep deprivation,” so drops in attention, mental agility, and problem-solving. Problem-solving was the “most heavily affected by sleep deprivation” (Das, Pratyush, Das, Maiti, & Krishna, 2022, p. 205). This work shows well that solid research designs are important, so the machine learning analytics of the data have relevance and impact.
Conclusion
Figure 5. Artificial Neural Network Illustration
Prasenjit Chatterjee, Parmanand Astya, Sudeshna Chakraborty, and Pooja, Editors of Machine Learning Algorithms for Engineering Applications: Future Trends and Research Directions (2022), highlights some practical applications of machine learning to solve human problems, by improving healthcare, strengthening network analysis, modeling hydrology, using wearable smart sensors, improving the food industry, enabling machine translation of language, strengthening tissue bio-engineering, drawing 3d point clouds from videos or image sequences, enhancing various dimensions in financial services, and studying the effects of sleep deprivation on human mental focus and performance.
About the Author
Shalin Hai-Jew works as an instructional designer / researcher at Kansas State University. Her email is shalin@ksu.edu.
| Previous page on path | Cover, page 13 of 22 | Next page on path |
Discussion of "Book review: Engineering life-enhancing solutions with machine learning"
Add your voice to this discussion.
Checking your signed in status ...