Practical Data Analytics for Innovation in Medicine: Building Real Predictive and Prescriptive Models in Personalized Healthcare and Medical Research Using AI, ML, and Related Technologies

Description

Practical Data Analytics for Innovation in Medicine: Building Real Predictive and Prescriptive Models in Personalized Healthcare and Medical Research Using AI, ML, and Related Technologies, Second Edition discusses the needs of healthcare and medicine in the 21st century, explaining how data analytics play an important and revolutionary role. With healthcare effectiveness and economics facing growing challenges, there is a rapidly emerging movement to fortify medical treatment and administration by tapping the predictive power of big data, such as predictive analytics, which can bolster patient care, reduce costs, and deliver greater efficiencies across a wide range of operational functions. Sections bring a historical perspective, highlight the importance of using predictive analytics to help solve health crisis such as the COVID-19 pandemic, provide access to practical step-by-step tutorials and case studies online, and use exercises based on real-world examples of successful predictive and prescriptive tools and systems. The final part of the book focuses on specific technical operations related to quality, cost-effective medical and nursing care delivery and administration brought by practical predictive analytics. Front Cover Practical Data Analytics for Innovation in Medicine Copyright Page Dedication Contents About the authors Foreword for the 2nd edition–John Halamka Foreword for the 1st edition by Thomas H. Davenport Foreword for the 1st edition by James Taylor Foreword for the 1st edition by John Halamka Preface and overview for the 2nd edition Preface to the 1st edition Modern medicine: an exercise in prediction and preparation Wasted costs in US healthcare systems References Acknowledgment Guest Chapter Author’s Listing Guest – Authors Endorsements and reviewer Blurbs—from the 1st edition References Instructions for using software for the tutorials—how to download from web pages—for the 2nd edition Prologue to Part I I. Historical perspective and the issues of concern for health care delivery in the 21st century 1 What we want to accomplish with this second edition of our first “Big Green Book” Prelude Purpose/summary First reasons for our writing this book Highlighted new material Descriptive statistics, data organization, and example Randomized controlled trials From observation to randomized controlled trials—eliminating bias Basic predictive analytics and example Example Research standards common to both traditional and predictive analytics Pandemic as related to research standards and accurate data Especially for the second edition Chapter conclusion Postscript References 2 History of predictive analytics in medicine and healthcare Prelude Outline Introduction Part I. Development of bodies of medical knowledge Earliest medical records in ancient cultures The oldest official medical documents Classification of medical practice among ancient and modern cultures Medical practice documents in major world cultures of Europe and the Middle East Egypt Mesopotamia Greece Medicine in Preclassical Greece Hippocrates and classical Greece Ancient Rome Galen Arabia Summary of royal medical documentation in ancient cultures Effects of the middle ages on medical documentation Rebirth of Interest in medical documentation during the renaissance The printing press The Protestant Reformation Erasmus Human anatomy Andreas Vesalius (1514–1564) William Harvey (1578–1657) Medical documentation after the enlightenment Medical case documentation The development of the National Library of Medicine Part II. Analytical decision systems in medicine and healthcare Computers and medical databases Early medical databases Medical literature databases National Library of Medicine list of online medical databases Other medical research databases Bills of Mortality in London, United Kingdom Best practice guidelines Guidelines of the American Academy of Neurology Medical records move into the digital world Healthcare data systems Postscript References 3 Bioinformatics* Prelude The rise of predictive analytics in healthcare Moving from reactive to proactive response in healthcare Medicine and big data An approach to predictive analytics projects The predictive analytics process in healthcare Process steps in Fig. 3.1 Step 1. Problem definition Step 2. Identify available data sources Step 3. Formulate a hypothesis Step 4. Data preprocessing Step 5. Data set design Step 6. Feature selection Step 7. Model building Step 8. Model evaluation Step 9. Model implementation Step 10. Validation of clinical utility Translational bioinformatics Clinical decision support systems Hybrid clinical decision support systems Consumer health informatics Patient-focused informatics Health literacy Consumer education Direct-to-consumer genetic testing Use of predictive analytics to avoid an undesirable future Consumer health kiosks Who uses the Internet? Nearly everybody Patient monitoring systems Applications for predictive analytics in intensive care unit patient monitoring systems Challenges of medical devices in the intensive care unit Public health informatics The major problem: lack of resources Social networks and the “Pulse” of public health Predictive analytics and prevention and disease and injury Biosurveillance Food-borne illness Medical imaging Clinical research informatics Intelligent search engines Personalized medicine Hospital optimization Challenges Data storage volumes Data privacy and security Standards and consistency of data Interpretability of models Evidence-based guidelines and adoption of PA models Portability of PA models Regulation of PA models Summary Postscript References Further reading 4 Data and process models in medical informatics Prelude Chapter purpose Introduction Systems for classification of diseases and mortality Bills of mortality The ICD system The OMOP common data model Reasons for OMOP The OMOP CDM provides a common data format OMOP CDM architecture is patient-centric Additional data processing operations necessary to serve the analysis of OMOP data The CRISP-DM processing model CRISP-DM phases How this chapter facilitates patient-centric healthcare Postscript References Further reading 5 Access to data for analytics—the “Biggest Issue” in medical and healthcare predictive analytics Prelude Size of data in our world: estimated digital universe now and in the future Convergence of healthcare and modern technologies Reasons why healthcare data is difficult to get and difficult to measure Multiple places where medical data are found Many different formats of medical data: structured and unstructured Another problem is inconsistent definitions Changing government regulatory requirements keep changing what data is taken and kept What are some of the benefits of using good data analytics in medical research and healthcare delivery? Conclusion of Chapter 5: the importance of health care data analytics Postscript References Further reading 6 Precision (personalized) medicine Preamble What is personalized/precision medicine? Personalized medicine versus precision medicine P4 medicine P5 to P10 medicine Precision medicine, genomics, and pharmacogenomics Differences among us Differences go beyond our body and into our environment Changes from birth to death Ancestry and disease Gene therapies It is not about just our genome Changing the definition of diseases Systems biology Efficacy of current methods—why we need personalized medicine Predictive analytics in personalized medicine The future: predictive and prescriptive medicine Application of predictive analytics and decisioning in predictive and prescriptive medicine The diversity of available healthcare data Diversity of data types available Phenotypic data Clinical information Real-time physiological data Imaging data Genomic data DNA—the center piece of heredity and bodily differences DNA replication and mutation Somatic mutations Germline mutations The personal genome project The Electronic Medical Records and Genomics network The Patient-Centered Outcomes Research Institute Transcriptomics data Epigenomics data Proteomic data Glycomic data Metabolomic data Metagenomic data Nutrigenomics data Behavioral measures data Socioeconomic status data Personal activity monitoring data Climatological data Environmental data All the other OMICs The future Challenges Challenge #1 Challenge #2 Challenge #3 Challenge #4 Challenge #5 Challenge #6 Challenge #7 Challenge #8 Challenge #9 Challenge #10 Challenge #11 Challenge #12 Challenge #13 Postscript References Further reading 7 Patient-directed healthcare Prelude Empowerment in patient-directed medicine Self-monitoring, N of 1 study Research questions The responsible patient Patients changing how medicine is practiced Patient empowerment versus compliance Collaboration between patients and the medical community Patient involvement Patient involvement in medical education Limitations of patient involvement Evidence supporting patient involvement Family-wise statistical errors Communication and trust Communication and trust during the pandemic Collaboration and limitations How patient-directed medicine works using predictive analytics Privacy concerns can hinder research Predictive analytics for patient-directed research Cultures and decisions Coordination of care and communication for patient-directed healthcare Communication skills in the medical setting Communication studies Barriers to productive communication Patients selecting their best models of care Medical homes The integrated healthcare delivery system model Comparison with accountable care organization Direct pay/direct care model Consumerism and advertising in patient-directed healthcare Advertising to patients Research studies related to advertising and consumerism Privacy of prescription data. Is it private? Patients diagnosing themselves amid targeted advertising Patients making use of technology and advertising for good or for bad Patient payment models and effects on self-directed healthcare Burden of healthcare—predicting the future Predicting life and death Misapplication of treatment increases costs Models of insurance—predicting the best for individuals Research assisting patients in self-education and decisions Patient self-responsibility: highlight on obesity Percent of obesity Distribution of obesity in the United States—costs and related diseases Cascading effects on sleep of obesity Obesity, cholesterol, statins, and patient-directed healthcare The need for N of 1 studies N of 1 study examples Data scientists could make a fortune—development of apps and artificial intelligence for phones and PC application Patient portals Alternatives and new models Medical tourism Where could it go wrong? Alternative screenings Self-diagnostic kits An alternative to traditional insurance Doctors striking out on their own Alternative ways of knowing about ourselves—genomic predictions Some concerns Predictive analytics for patient decision-making Connectivity Controlling some diseases by searching research on one’s own Portals, evidence medicine, and gold standards in predictive analytics Patientsite at Beth Israel Cleveland clinic Body computing Diagnostic apps Chapter conclusion Postscript References 8 Regulatory measures—agencies, and data issues in medicine and healthcare Prelude Introduction What is an electronic medical records? Five of the best open source electronic medical records systems for medical practices Rise of the international classification of disease Six Sigma Quality control Lean concepts for healthcare: the lean hospital as a methodology of Six Sigma Root cause analysis Henry Ford Hospitals and Virginia Mason Hospital Postscript References Further reading 9 Predictive analytics with multiomics data Prelude Introduction to multiomics Genomics Multiomics Multiomics systems biology Basic analytics operations in multiomics Multiomics data integration Multiomics data preparation Methodological bias Unrepresentative negatives Imbalance of data sets with rare target variables Data preparation issues specific to particular omics data sets Microarray data Gene sequencing data Mass spectrometer data Peak extraction from the chromatogram Peak detection Peak merging and annotation Peak alignment Peak registration Peak filling Analysis methods Statistical analysis methods Machine learning methods Data conditioning Normalization Data set balancing Data preprocessing tools in multiomics Multiomics analytical methods Open source tools for multiomics analytics Machine learning tools in multiomics analytics Filter methods Wrapper methods Embedded methods Deep learning for genome and epigenome analysis Focus on metabolomics Prediction of pancreatic and lung cancer from metabolomics data Postscript References Further reading 10 Artificial intelligence and genomics Prelude How do we enable the clinical application of artificial intelligence in genomics? Genomics fast moving field—and now ready for artificial intelligence to have an impact Need to open existing large datasets to more researchers Successful artificial intelligence models will be ones that use smaller and manageable portions of the human genome Polygenic risk scores Artificial intelligence models cannot replace but must augment physicians diagnosis and treatment decisions Governance—balance between rapid approval of models and ensuring no human harm EHR and integration of artificial intelligence into clinical workflows What would an artificial intelligence and genomics integration look like? Real-world examples of artificial intelligence and genomics modeling systems emerging in 2022 Conclusions Postscript References Further reading Prologue to Part II II. Practical step-by-step tutorials and case studies Prologue to Part III III. Practical application examples 11 Glaucoma (eye disease): a real case study; with suggested predictive analytic modeling for identifying an individual pat... Prelude Why this chapter in this book? How serious is glaucoma? Why do we need to watch for it? What is a normal eye pressure? Characteristics of glaucoma disease Risk factors and treatment Basic anatomy of the eye and relation of physical structure to glaucoma disease What is glaucoma? What is the normal pressure (IOP) in the eye? What causes a rise in intraocular pressure above the norm of 10–21? Pathophysiology of glaucoma Diagnosis of glaucoma Illustrations/photo of eye “Minimally invasive” surgeries can be invasive Invasive surgical treatments Trabeculectomy Trabeculotomy What does the XEN-gel stint look like? What is its size? Ahmed valve shunt. What does the Ahmed valve shunt look like? Long-term results of using Ahmed valve shunts for glaucoma Fluid flow in the two main types of glaucoma Open angle Closed angle Photography of eye—looking at fundus in the diagnosis of glaucoma Case study: my (Gary’s) glaucoma progression (from about 2010 to 2022) Self-monitoring intraocular pressure by the patient for more accurate DX and treatment decisions i-CARE home device for patient home monitoring of intraocular pressure values As others are stating My invasive surgery—2021—XEN-gel shunt and later Ahmed valve shunt Increased night-time urination frequency was an unpleasant side-effect of my using steroid eyedrops Is increase in “urination frequency” a common side effect of use of “steroids in eye drops”? Suggested absorbsion pathway of Loetmax SM; Helping to determine best treatment Predictive analytic modeling possibilities Even visual field tests can now be automated with artificial intelligence—machine learning methods Using STATISTICA statistical and predictive analytic software to visualize patient Gary’s IOP data DOSE OF “Generic-COSOPT” (=Dorzolamide–Timolol)—is three times a day OK? Dosing Future possible treatments for glaucoma FINAL IOP levels for Gary upon finding “optimum mix of steroid and IOP eye drops” Postscript References Further reading 12 Using data science algorithms in predicting ICU patient urine output in response to diuretics to aid clinicians and heal... Prelude Introduction Outputs and conclusion from a literature review The data used Source of data Data demographics Technology used Data science author Data science operations Algorithm outputs and decisions Algorithm version 1 Introduction Algorithm selection Selection criteria—algorithm Steps A and B Selection criteria—algorithm Step C Selection criteria—algorithm Step D Short description of each algorithm Regression in general and classification and regression trees Boosted trees Random forests Algorithm results Summary of R-squared values projected to predict UO at hour 6, 12, and 24 How is R-squared used and interpreted Detailed prediction results Hour 1 results for predicting hour 6 Hour 3 results for predicting hour 6 Hour 6 results or predicting hour 12 Hour 8 results or predicting hour 12 Hour 12 results or predicting hour 24 Hour 18 results or predicting hour 24 Algorithm step C: calculation of targets Algorithm step D: predicting if patients would reach targets (yes/no)—target at 75th percentile Predicting a yes/no for hour 24 at hour 6, 12, and 24 Hour 6 Hour 12 Hour 24+ Algorithm step D: predicting if patients would reach targets (yes/no)—target at 95th percentile Hour 6, 12, and 24+ Hour 6 Hour 12 Hour 24+ Algorithm accuracy, sensitivity and specificity for the classification trees above ROC curve for algorithm Step D: predicting if patients would reach targets (yes/no)—target at 75th percentile ROC curve for algorithm step D: predicting if patients would reach targets (yes/no)—target at 95th percentile Algorithm version 2 Introduction Algorithm selection Short description of each algorithm Algorithm results Summary of R-squared values projected to predict UO at hour 6, 12, and 24 How is R-squared used and interpreted Detailed prediction results Hour 1 results for predicting hour 6 Hour 3 results for predicting hour 6 Hour 6 results or predicting hour 12 Hour 8 results or predicting hour 12 Hour 12 results or predicting hour 24 Hour 18 results or predicting hour 24 Predictor importance for the above Algorithm step C: calculation of targets Algorithm step D: predicting if patients would reach targets (yes/no)—target at 75th percentile Predicting a yes/no for hour 24 at hour 6, 12, and 24 Algorithm accuracy, sensitivity, and specificity for the classification trees above ROC curve for algorithm step D: predicting if patients would reach targets (yes/no)—target at 75th percentile ROC curve for algorithm step D: predicting if patients would reach targets (yes/no)—target at 95th percentile Algorithm version 3 Introduction Algorithm selection Short description of each algorithm Algorithm results Summary of R-squared values projected to predict UO at hour 6, 12, and 24 How is R-squared used and interpreted Algorithm step C: calculation of targets Algorithm step D: predicting if patients would reach targets (yes/no)—target at 75th percentile Predicting a yes/no for hour 24 at hour 6, 12, and 24 Hour 6 Hour 12 Hour 24+ Algorithm step D: predicting if patients would reach targets (yes/no)—target at 95th percentile Hour 6, 12, and 24+ Hour 6 Hour 12 Hour 24+ Algorithm accuracy, sensitivity and specificity for the classification trees above ROC curve for algorithm step D: predicting if patients would reach targets (yes/no)—target at 75th percentile ROC curve for algorithm step D: predicting if patients would reach targets (yes/no)—target at 95th percentile The champion algorithms Further research not published here—a champion emerges The conclusions on our champion algorithm Examples to illustrate model performance for actual patients Conclusions and further recommendations Conclusions Recommendations Postscript Further reading 13 Prediction tool development: creation and adoption of robust predictive model metrics at the bedside for greatly benefit... Prelude Author’s note Rationale Exploratory data analysis for health data Differences between traditional analysis and exploratory data analysis driven analysis Prediction tool development The need for a prediction tool for premature infants at risk of bronchopulmonary dysplasia Modeling output from the estimator to improve prediction Methods Obtaining and processing data Data tidying Using R Shiny for efficient data input and visualization After obtaining the finalized clean data Training and testing dataset TRIPOD guidelines highlights Code examples and tutorial Data cleaning and TidyR examples Initializing an R Shiny web app The user interface The server Loading and saving onto a SQL database Showing and interacting with data Reactive expressions to account for computationally or temporally expensive algorithms Conclusion Appendix Download links Versions of software and packages Postscript References Further reading 14 Modeling precancerous colon polyps with OMOP data Prelude Chapter purpose Introduction The University of California, Irvine Colonoscopy Quality Database The UCI Colon Polyp Project Previous colon cancer risk screening and predictive modeling programs OMOP data Caveat Modeling objective Methods Major tasks of data preparation of OMOP data for modeling Data access The modeling tool Data integration Target variable definition Data type changes Data quality assessment and resolution Data exclusions Aggregation to the patient level Unique code determination Text mining frequency analysis Manual variable derivation Derivation of one-hot (binary) variables Feature selection process The “short-list” Methods of feature selection Variable filtering Wrapper methods Data conditioning Balancing the data set Unrepresentative negatives Positive unlabeled learning Modeling Modeling algorithms Cross-validation Ensemble modeling Results and discussion Model evaluation Prediction accuracies Receiver operator characteristic curve Other important aspects of the trained model Important predictor variables Emergent properties Automation of data preparation for medical informatics? Conclusions How this chapter facilitates patient-centric medical health care Postscript References Further reading 15 Prediction of pancreatic and lung cancer from metabolomics data Prelude Purpose of this chapter Introduction Cancer deaths in the United States Cancer metabolites Methods The modeling process Results Model accuracy Specific models for lung cancer and pancreatic cancer Discussion Implications of this case study for future medical diagnosis Conclusions How this chapter facilitates patient-centric healthcare Postscript References 16 Covid-19 descriptive analytics visualization of pandemic and hospitalization data Preamble Introduction 3 KNIME workflow data streams Preparatory steps for using this tutorial General introduction to KNIME Data access—the file reader node Data understanding Country selection Visualization data stream Using the workflow for another country How this chapter facilitates patient-centric healthcare Postscript Further reading 17 Disseminated intravascular coagulation predictive analytics with pediatric ICU admissions Prelude Introduction Background (from first edition) The example Data files First week of analysis Data mining recipes using statistica Data imputation Using the 11,459 imputed file—training data Training data (11,569 imputed) continued A problem Randomly separating the data and new data mining recipe Final analysis—a return to the past Conclusion—personal ending thoughts Postscript References Prologue to Part IV IV. Advanced topics in administration and delivery of health care including practical predictive analytics for medicine... 18 Challenges for healthcare administration and delivery: integrating predictive and prescriptive modeling into personalize... Prelude Introduction to challenges in healthcare delivery Challenge #1 A new infrastructure must be created for healthcare Challenge #2 Who will pay for gene sequencing of every individual for healthcare purposes? Challenge #3 Who will regulate predictive models that that directly affect patient care? How will these models be tested for clinical use? Challenge #4 What effects will predictions of health outcomes have on an individual’s mental health and general overall daily anxiety le... Challenge #5 Does the ability to predict someone’s health outcome change their behavior? Challenge #6 With the possibility of building clinical risk prediction models also comes the possibility of legal liability Challenge #7 Many of the biological (body) specimens that are required for full “omics analysis” are not easily obtained Challenge #8 The technology required for personalized medicine is certainly not perfect, and many refinements and further developments a... Challenge #9 It is likely that the effective implementation of predictive–precision medicine models in medicine will require more suppor... Challenge #10 The scientific disciplines covered by personalized–predictive medicine are numerous Challenge #11 Accurately capturing data from the medical record for use in predictive models Postscript References Further reading 19 Challenges of medical research in incorporating modern data analytics in studies Prelude Introduction—challenges to medical researchers Trends that we might want to know about Automation and machine learning (AutoML) Blockchain Conversational artificial intelligence Digital twins Medical competitions Conclusion Postscript References Further reading 20 The nature of insight from data and implications for automated decisioning: predictive and prescriptive models, decision... Prelude Overview The purpose of this chapter The nature of insight and expertise Procedural and declarative knowledge Nonconscious acquisition of knowledge Conclusion: expertise and the application of pattern recognition methods Statistical analysis versus pattern recognition Fitting a priori models Pattern recognition: data are the model The data are the model Pattern recognition in artificial intelligence/machine learning: general approximators Pattern recognition and declarative knowledge: interpretability of results Explainability of artificial intelligence/machine learning models Global and local explainability Statistical models, and reason scores for linear models What-if, and reason scores as derivatives Explainability of nonlinear models, artificial intelligence/machine learning models Local interpretable model-agnostic explanations Shapley additive exPlanations Comparing local interpretable model-agnostic explanations and Shapley additive explanations Caution: inverse predictions can be very risky Inverse prediction Correlation is not necessarily causation Lack of evidence at the specific point in the input space Optimization of inputs to achieve a desired output Naive explanations Summary Postscript References 21 Model management and ModelOps: managing an artificial intelligence-driven enterprise Prelude Introduction The model building/authoring life cycle Overview: managing the life cycles for thousands of models Types of analytic models Managing the risks of analytics, artificial intelligence Do-no-harm Industry certifications ModelOps scope ModelOps details: managing model pipelines and reusable steps The tools and languages of artificial intelligence/machine learning Reusable steps, building intellectual property Avoiding analytic silos, big code Siloed data science Managing input data and output sinks Validating data Validating outputs Managing model life cycles Proposing a new model, updating a model Deployment environments Roles, personas, approvals Granular permissions Model monitoring Beyond model accuracy Differential accuracy Linking model predictions to value Monitoring for model drift, shift Concept drift Population stability and drift Data drift Monitoring risks Unexpected consequences Anticipating risk Efficiency, agility, elasticity, and technology Cloud architecture Private and public cloud On-prem cloud solutions Cloud architecture and containerization Managing models for data-at-rest and data-in-motion Addressing real-time scoring requirements Responding to a dynamically changing world Conclusion Postscript References Further reading 22 The forecasts for advances in predictive and prescriptive analytics and related technologies for the year 2022 and beyond Prelude Section I: specific technological trends predicted for 2022–2023+ What is predictive analytics, and what are the most frequently used methods (or algorithms) in predictive analytics? What is prescriptive analytics, and what is an example of prescriptive analytics? Part I—healthcare: what trends can we expect in the year 2022 and beyond? What do these three things mean? Part 2—In general: PA and business intelligence trends for 2022 TOP 10 analytics and business intelligence trends for 2022 Key artificial intelligence and data analytics trends for 2022 and beyond Section II: overriding philosophies which will guide trends over the next 10 years Postscript References 23 Sampling and data analysis: variability in data may be a better predictor than exact data points with many kinds of Medi... Prelude Sampling and data analysis issues Purpose summary of this chapter One issue—electronic health record and specific measures taken on patients Pulse oximetry data measurements, as an example Introduction Objective Methods Results Discussion Conclusion on Pulse Oximetry Example Eye-intraocular pressure measurements: a personal example by one of the authors to illustrate the problem of when and how d... Example of comparison of Goldman with i-CARE HOME intraocular pressure readings In conclusion Types of data analysis that may be helpful in solving the types of issues presented in this chapter Reliability of inputs determines the validity of models However, it gets more complicated But then, it gets even more complicated Clinical Dx and treatment needed changes for true patient-centered care Postscript References Further reading 24 Analytics architectures for the 21st century Prelude Introduction Purpose/summary Organizational design for success Some say it starts with data, it doesn’t Organizational alignment Framework for trustworthy and ethical AI and analytics Data design for success Why is data so important? The potential of data is insight and action Data and analytics literacy are requirements to successful programs Brief considerations in data architecture Processes, systems, and data Data volume Data variety Data velocity Data value Data veracity Connecting and moving data—data in motion Application programming interfaces and management Microservices Streaming data Data stores and limitations of the enterprise data warehouses Timeliness of data Data additions are slow and difficult Some data will never be available in the enterprise data warehouses Ancillary and cloud data Data virtualization A platform with searchable data and rich metadata A collaboration tool for functional areas and users A pathway for new systems and system migration An IT tool for rapid prototyping A system for enhanced security of data Data governance and data management Goals of data governance Technology to support data management and governance Data management Master data Reference data Metadata Data quality Security Data governance and data management summary Analytics design for success Technology to create analytics Data discovery and acquisition Exploratory data analysis Data preparation Feature engineering Model build and selection Model evaluation and testing Model deployment Model monitoring Legality and ethical use of data Technology to communicate and act upon analytics Model Ops Conclusion Postscript References 25 Predictive models versus prescriptive models; causal inference and Bayesian networks Prelude Introduction Classification of AI and ML models in medicine Descriptive analytics Diagnostics analytics Predictive analytics Prescriptive analytics Process optimization Causation—the most misunderstood concept in data science today Some basic assumptions for predictive modeling Some basic assumptions for prescriptive modeling Using a predictive model for prescription purposes Three considerations for transforming a predictive model into a prescriptive model Assume causation based on well-documented scientific study Design of experiments and randomized clinical trials—the traditional gold standard Causal inference methods that may include Bayesian networks Some important notes on observational studies Causal inference and why it is important Bridging the causal models to statistical models—causal inference Bayesian networks Benefits of Bayesian networks Drawbacks of Bayesian networks Causal inference and the do-calculus A summary example of causal modeling Conclusion Postscript References Further reading 26 The future: 21st century healthcare and wellness in the digital age* Prelude Overview Background and need for change Comparative effectiveness research and heterogeneous treatment effect research New technology and 21st century healthcare: health startup firms Building the “Star Trek” tricorder We wrote this back in 2014 for the first edition of this book Well did this all happen as predicted? Not quite Listing of other e-items in this “outside of healthcare facilities” category but within at least the partial control of pat... Examples of wearable devices that are working for people today Atrial fibrillation wearable watch sensors Augmented reality app EKG—home monitoring digital system Eye pressure (IOP) home measurement devices Nonautomatic vital health signal measuring devices Blood pressure devices Oxygen level home monitors Trends and expectations for the future of health IT and analytics Bottom-Up “small-sized” but working individually controlled data gathering and instant analytics output systems Where will the next innovations in medicine come from? N-of-1 studies—the future for person-centered healthcare Styles of thinking—how brain laterality affects innovation in healthcare Final concluding statements How much should we listen to algorithms?—Should machines make the decisions? Genomics and AI will start exploding in 2022 and subsequent years, and thus we need to be prepared Patient-centered (precision) health for the future Postscript References Further reading Appendix A Modeling new COVID-19 deaths Introduction Processing steps in this workflow Summary Reference Index Back Cover