Chapter 6 - Data Science and AI

data science

• the science of learning from data

  • studies methods involved in analysis n processing of data n tech to improve methods in evidence-based manner

  • gain new insight

data analytics

• use data n analysis to to make fact based decisions

• for planning, management, measurement, learning

machine learning (ML)

• gives computer ability to learn without being explicitly programmed

  • example data or past experience

• falls under ai

machine learning (ML) derived from

• math / stats

  • learn relationships from data

• comp sci

  • efficient algorithm, esp. with large amts of data

deep learning

• machine learning that use neural networks (small processing units) with deep layers

• most successful with image analysis

artificial intelligence (AI)

• area of comp sci focused on developing info systems n algorithms that can perform tasks associated with human intelligence

• initially for engineering knowledge representations n algorithm

  • now given way to ML

generative ai

• creates new content based on data

Big data 4 Vs

• volume

  • amt of data increasing

• velocity

  • quickly generated

• variety

  • diff types

• variability

  • from trustable sources

  • data not always consistent

data mining

• processing n modeling data to discover previously unknown patterns or relationships

  • patterns alr exist in data

text mining

• applying data mining to unstructured textual data

  • look for patterns within clinical notes

natural language processing (NLP)

• help computer understand human lang. (written text n speech)

data provenance

• origin n trustworthiness of data

business intelligence

• use of data to get timely, valuable insights into business n clinical data

precision medicine

• was formerly called personalized or computational med

• 2 patients get same diagnosis, but diff treatments

metadata

• data abt data

• data context

  • what the data represents

data visualization

• use visual methods to “tell story” abt data

data wrangling

• prep n process data into format used for analytics, learning, visuals, etc

use cases for big data

• high cost patients - intervene early

• readmission - prevent

• triage - appropriate level of care

• decompensation - patient condition worsen

• adverse events - awareness

• treatment optimization - esp for diseases affecting multiple organ systems

requirements for data analytics in healthcare

• infrastructure

  • stakeholder engagement

  • human subject research protection

  • protect patient privacy

  • data assurance n quality

  • interoperability of health info systems

  • transparency

  • sustainability

universal data architecture

• where info organized from diff places

amt of EHR data varies bc

• patients get care at diff places

• sicker patients more data

• data completeness varies

  • hm data available in record

• absent / discordant (inconsistent) documentation

• rare diseases take longer to diagnose

informed presence bias

• people present in EHR may be diff from people who aren’t receiving care

  • more EHR ≠ more disease

issues w operational EHR data

• inaccurate

• incomplete

• transformed in ways that undermine meaning

• unrecoverable

• unknown provenance

• insufficient granularity

• incompatible w research protocols

approaches to ML

• supervised

• unsupervised

• semi-supervised

• reinforcement learning

ML supervised learning

• learn to predict known output

  • training data (teaches)

  • evaluated on test data

    • avoid over fitting

ML unsupervised learning

• find naturally occurring patterns or groupings within data

ML semi-supervised learning

• mix of supervised n unsupervised , labeled n unlabeled inputs

  • algorithm find structure n patterns on their own with help from labeled inputs

reinforcement learning

• learns from ongoing data n results

biomedical applications of ML

• imaging

• clinical prediction

• biological processes

• assisting humans

sensitivity

• detect true positives for a disease

  • ppl who have it

specificity

• can declare disease not present

  • true-negatives

when implementing AI in real world

• data variability across institutions

• impact on clinical workflows - esp. nurses

• short overviews of purpose n potential harms

• stewardship of algorithm for efficacy n safety

tools for ML n AI

• R

• python

• jupyter notebook

R tool

• for stat computing n graphics

  • “tidy” data

python tool

• easy use n lang

• popular for data sci n ML

jupyter notebooks

• local web applications contain live code, equations, figures, interactive apps, markdown text

non-programming packages

• orange

• RapidMiner

challenges for ML n AI

• ethics n bias

• explainability

• reproducibility

• regulation n liability

explainability

• explain outputs, esp from neural networks

black box med

• ai gives answer, may be difficult to understand how it got the ans

reproducibility

• make sure algorithm can be tested / validated

  • work across diff datasets

regulation n liability

• who responsible for use or non-use of ai

• how ai regulated due to changes over time

• US FDA develop guidelines for software as medical device

• AMIA position paper on adaptive clinical decision support

ML n AI evaluation

• use evidence-based med (EBM) to access prediction/diagnosis models

• clinical implementation typically assessed using randomized controlled trials (RCTs)

studies evaluating ML n AI

• have

  • meaningful endpoints (benefits)

  • appropriate benchmarks (real world use)

  • interoperable n generalizable (transportable to other settings n systems)

  • specified interventions

  • audit mechanisms (monitored after implementation)

  • promise n protection (legal n ethical monitoring)

• number needed to benefit

  • account for benefit of use of tool n clinical impact

• critical appraisal of models must evaluate model development n outcomes

• most studies don’t use standard methods

ML n AI impact clinical practice

• physicians n ML adapt to each other

• ai wont replace radiologists, but radiologists using ai will replace ones who don’t