Winda

MIT 6.S897: Machine Learning for Healthcare

Contents · 50

  1. 01-What Makes Healthcare Unique?The history of AI and machine learning in healthcare, the growth of electronic health records and multimodal data, and opportunities and challenges for applying ML to clinical care.
  2. 01-What Makes Healthcare Unique?Introduces machine learning’s potential to address healthcare costs, missed diagnoses, and medical errors, with examples from AI’s history and an overview of the course’s focus on applying ML in healthcare.
  3. 02-Overview of Clinical CareIntroduces health care goals and measures of mortality, disability, morbidity, and quality of life, then surveys diagnosis, prognosis, treatment, prevention, and care processes.
  4. 02-Overview of Clinical CareHealth care’s goals, definitions of health, life expectancy, the role of healers, medical practice, and the motivations behind health care spending.
  5. 03-Deep Dive Into Clinical DataCovers clinical data types, demographic and heart-rate analyses, mortality modeling, and prescription standards in MIMIC-III.
  6. 03-Deep Dive Into Clinical DataMIMIC clinical data examples highlight measurement artifacts from system changes, age de-identification, and neonatal ICU coverage.
  7. 04-Risk Stratification, Part 1Introduces patient risk stratification and its evaluation, then develops a machine-learning approach to early type 2 diabetes detection using longitudinal administrative and clinical data.
  8. 04-Risk Stratification, Part 1Introduces clinical risk stratification, its goals and differences from diagnosis, with examples including early diabetes detection and infant morbidity prediction.
  9. 05-Risk Stratification, Part 2Covers label derivation, risk-model evaluation, non-stationarity, intervention-tainted outcomes, and an introduction to survival modeling.
  10. 05-Risk Stratification, Part 2Deriving outcome labels from health records, evaluating risk models, and handling bias and censoring in survival analysis.
  11. 06-Physiological Time-SeriesSurvival modeling and physiological time-series, including neonatal ICU monitoring, artifact detection, switching state-space models, and atrial fibrillation detection.
  12. 06-Physiological Time-SeriesRisk stratification and survival modeling with censored data, followed by physiological time-series modeling for ICU monitoring and EKG diagnosis.
  13. 07-Natural Language Processing (NLP), Part 1Clinical NLP for extracting diagnoses, medications, laboratory findings, and other entities from medical records, including negation handling, cohort identification, and model portability.
  14. 07-Natural Language Processing (NLP), Part 1Why clinical text is valuable, limitations of billing codes for identifying disease, and methods for extracting clinical concepts from narratives.
  15. 08-Natural Language Processing (NLP), Part 2Clinical phenotyping from emergency department records using anchor-and-learn methods, with attention to NLP features, context, parsing ambiguity, and entity-level features.
  16. 08-Natural Language Processing (NLP), Part 2Clinical NLP methods for expanding phenotype vocabularies from anchor terms, evaluating cohort selection, and using context to distinguish clinical concepts from identifying information.
  17. 09-Translating Technology Into the ClinicExamines challenges in translating health technology into clinical practice, contrasting Watson’s limits with benefits of computerized physician order entry.
  18. 09-Translating Technology Into the ClinicAI hype cycles, venture capital, expert systems, and the risks of unmet expectations and AI winter.
  19. 10-Application of Machine Learning to Cardiac ImagingCovers cardiac structure and function, major heart diseases, and imaging modalities used in diagnosis and medical decision-making.
  20. 10-Application of Machine Learning to Cardiac ImagingCardiovascular medicine, cardiac diagnostics, and the use of machine learning and computer vision to automate echocardiogram interpretation and reveal biological insights.
  21. 11-Differential DiagnosisDiagnostic reasoning for cardiovascular dynamics, including differential diagnosis models, Bayesian testing, test performance, and utility-based decision analysis.
  22. 11-Differential DiagnosisIntroduces differential diagnosis and diagnostic reasoning, including clinical signs and symptoms, physiological models, triage flowcharts, disease associations, and probabilistic and utility-based methods.
  23. 12-Machine Learning for PathologyMachine learning and deep learning in pathology, including diagnostic variability, digital whole-slide image analysis, and breast cancer metastasis detection.
  24. 12-Machine Learning for PathologyComputational pathology and AI applications to tissue-image analysis, cancer diagnosis, disease subtyping, and treatment decisions.
  25. 13-Machine Learning for MammographyExplains mammogram cancer triage, dataset collection, model design and training, image-level score aggregation, safe triage thresholds, and probability calibration.
  26. 13-Machine Learning for MammographyMammogram cancer detection and triage, including screening workflow, dataset design, modeling challenges, and analysis.
  27. 14-Causal Inference, Part 1Introduces potential outcomes, treatment effects, counterfactuals, and confounding in healthcare decisions.
  28. 14-Causal Inference, Part 1Why causal reasoning is essential in healthcare, with examples from diabetes prevention, cancer prognosis, and treatment decisions.
  29. 15-Causal Inference, Part 2Covers covariate adjustment, matching, and propensity score weighting for estimating treatment effects, with linear and nonlinear models.
  30. 15-Causal Inference, Part 2Introduces potential outcomes, conditional and population average treatment effects, and applies causal inference concepts to COVID-19 questions, including the limits of observational data and assumptions about social interactions.
  31. 16-Reinforcement Learning, Part 1Introduces sequential treatment policies and reinforcement learning for clinical decision-making, using sepsis care to explain decision processes, MDPs, and contextual bandits.
  32. 16-Reinforcement Learning, Part 1Sequential treatment decisions for sepsis management, causal effects, and policies that choose actions based on patient state.
  33. 17-Reinforcement Learning, Part 2Dynamic treatment strategies, g-methods, and the parametric g-formula, with a prostate cancer case study of physical activity and survival.
  34. 17-Reinforcement Learning, Part 2Policy evaluation in reinforcement learning, including limits of average reward objectives and alternatives that account for worst-case outcomes.
  35. 18-Disease Progression Modeling and Subtyping, Part 1Alzheimer’s disease progression prediction with multi-task learning, sparse temporal models, and unsupervised disease subtyping via clustering.
  36. 18-Disease Progression Modeling and Subtyping, Part 1Introduces disease progression modeling and staging, using multiple myeloma to explore patient trajectories, prognosis, progression timing, and treatment effects.
  37. 19-Disease Progression Modeling and Subtyping, Part 2Methods for modeling disease progression from cross-sectional patient data, including latent disease staging, subtyping, Markov processes, noisy-OR models, and clinical anchors.
  38. 19-Disease Progression Modeling and Subtyping, Part 2Introduces cross-sectional disease staging from biomarker measurements, including progression assumptions, ambiguity in stage direction, and using age or time to death to resolve it.
  39. 20-Precision MedicinePrecision medicine uses clinical, temporal, imaging, and molecular data to classify patients, identify disease subtypes, and guide treatment. Covers patient trajectories, machine-learning challenges, and basic genetics.
  40. 20-Precision MedicineExplores precision medicine through disease subtyping, genetic and clinical data integration, large-scale patient cohorts, and the promise and challenges of genomic research.
  41. 21-Automating Clinical Work FlowsHow guidelines and protocols reduce variation in medical care, with examples from cholesterol management and risk assessment.
  42. 21-Automating Clinical Work FlowsHow clinical guidelines and protocols reduce variation in medical care, weighing consistency against improvements in average performance.
  43. 22-Regulation of Machine Learning and Artificial Intelligence in the USUS oversight of AI/ML health products, including FDA software and cybersecurity policies, public comments, and IRB involvement in digital research.
  44. 22-Regulation of Machine Learning and Artificial Intelligence in the USFDA regulation of AI and digital health products, including public comment processes, and institutional review board oversight, ethics, and data security.
  45. 23-FairnessIntroduces machine-learning fairness through criminal justice examples, technical biases in data and models, and the challenges of defining fair decisions.
  46. 23-FairnessThe talk examines AI decision-making in law, especially algorithmic fairness, bail and sentencing, judicial judgment, and the risks of automated systems.
  47. 24-Robustness to Dataset ShiftExamines dataset shift in healthcare, including temporal and institutional changes, population-level deployment checks, transfer learning, and adapting clinical prediction models across EHR systems. It also introduces adversarial perturbations and defenses.
  48. 24-Robustness to Dataset ShiftData set shift and covariate shift, including adversarial perturbations and how changes in data distributions can affect model performance.
  49. 25-InterpretabilityCovers model complexity, trust and generalizability, and explanation methods such as MYCIN and LIME, including qualities that make explanations useful.
  50. 25-InterpretabilityIntroduces model interpretability, human cognitive limits on explanations, model complexity and overfitting, and simplified “just-so” stories as a way to explain complex systems.