ai and machine learning in healthcare

47
MTREC Listen@Lunch AI and Machine Learning in Healthcare Welcome to the presentation. To prevent interruptions and to promote a learning environment, users have been muted and will not be able to unmute themselves until the Q&A at the end of the presentation. Please standby and the conference will begin shortly. Feel free to ask questions in the chat. We will respond during the Q&A. This presentation is being recorded and links will be shared after the conference.

Upload: others

Post on 12-May-2022

3 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: AI and Machine Learning in Healthcare

MTREC Listen@LunchAI and Machine Learning in

Healthcare• Welcome to the presentation.

• To prevent interruptions and to promote a learning environment, users have been muted and will not be able to unmute themselves until the Q&A at the end of the presentation.

• Please standby and the conference will begin shortly.

• Feel free to ask questions in the chat. We will respond during the Q&A.

• This presentation is being recorded and links will be shared after the conference.

Page 2: AI and Machine Learning in Healthcare

AI and Machine

Learning in

Healthcare

Nov 2020

Dr. Scott Moen

Page 3: AI and Machine Learning in Healthcare

Outline

What is artificial

intelligence?

What is machine learning?

What is deep learning?

General AI and meta AI

AI in medical coding

AI in clinical research

AI in diagnostics

AI in clinical infrastructure

Implementation details

Page 4: AI and Machine Learning in Healthcare

What is

artificial

intelligence?

A computer program

that is perceived as

human intelligence.

AI ML DL

Page 5: AI and Machine Learning in Healthcare

What is

machine

learning?

A computer program that

derives patterns from data,

instead of relying on explicit

instructions.

AI ML DL

Page 6: AI and Machine Learning in Healthcare

How does ML differ from standardprogramming?

y=mx+bMachine LearningTraditional Programming

Input data

Programmerdefinedprocess

Result

M =2

X = 3

B = 4

def line(M,X,B):

y= M*X+B

return y

Y=10

Training data Input Data

M =2

X = 3

B = 4M =2

X =5

B = 2

???

Fit Model

Training

Results

M =1

X =1

B =4

Y=10

M =2

X = 3

B = 4

Run Model

Inference

Result Y=10

Y = model

.predict(M,X,B)

Y=20

Y=5

Page 7: AI and Machine Learning in Healthcare

Some of the important types of ML and common applications

Unsupervised Learning – the training data has no labels• K-means clustering - an app that groups people based on

personality preferences (music, food choices, etc.}

• Kernel Density Estimation – an app that uses fishing vesselhauls and GPS data to predict the locations of schools offish.

Page 8: AI and Machine Learning in Healthcare

Some of the important types of ML and common applications

Semi-supervised Learning – some of the training data has labels

▪ An app that predicts crop infestation based on pictures of the damage. The algorithm will cluster the pictures based on feature similarity and then name it given the available labels in that cluster.

Page 9: AI and Machine Learning in Healthcare

Some of the important types of ML and common applications

Supervised Learning – the training data is labeled• Classification – an app that predicts the type of fish from a

photo and gives a percent for each category.

• Regression - an app that predicts a house price based on existing data.

Page 10: AI and Machine Learning in Healthcare

Some of the important types of ML and common applications

Reinforcement learning – the algorithm is rewarded/punished for actions

▪ A program plays Pac-Man and is given 1 pt. for every dot eaten and loses 5 pts for each time a ghost kills it. Additionally, the program is taught how to move up/down/left/right. It iterates over and over until a high score threshold is reached.

Page 11: AI and Machine Learning in Healthcare

ML strengths/weaknesses

• User doesn’t have to explicitly define pattern. Especially useful when the data is immense.

• User doesn’t need to know all the possible outcomes.

• The algorithm doesn’t have human biases.

ML strengths

• Garbage in-garbage out

• No common sense

• Algorithm only good for one purpose

ML weaknesses

Page 12: AI and Machine Learning in Healthcare

What is deep

learning?

Deep learning uses networks

that vaguely resemble human

neurons.DOI: 10.1007/978-3-030-00937-3_7

AI ML DL

Page 13: AI and Machine Learning in Healthcare

What differs from standard programming?

from keras.models import Sequential

from keras.layers import Dense

#create model

model = Sequential()

#get number of columns in training data

n_cols = train_X.shape[1]

#add model layers

model.add(Dense(4, activation='relu’,

input_shape=(n_cols,)))

model.add(Dense(5, activation='relu'))

model.add(Dense(1))

Input Layer Hidden Layer Output Layer

Page 14: AI and Machine Learning in Healthcare

Supervised – the training data is labeled

▪ Multilayer Perceptrons – great for basic classification and regression problems; excel sheet type of problems.

▪ Convolutional Neural Networks – the modern version of the perceptron, can handle very complicated data like self-driving cars.

Some of the important types of DL and common applications

Page 15: AI and Machine Learning in Healthcare

Unsupervised – the training data has no labels, learns by clustering

• Autoencoder – great for reducing very complex systems, like an app that looks at all the reviews of a specific movie and finds specific characteristics that certain people like or don’t like.

Some of the important types of DL and common applications

Page 16: AI and Machine Learning in Healthcare

General AI and meta AI

General AI

▪ An AI capable of handling multiple types of data.

▪ GPT-3

▪ IBM Watson

Page 17: AI and Machine Learning in Healthcare

General AI and meta AI

Meta-learning AI

▪ Combines multiple types of AI to get best answer or can be used to optimize hyperparameters of other machine learning models.

Page 18: AI and Machine Learning in Healthcare

3 General Applications of AI

Computer Vision

▪ Algorithms that automate human visual tasks by finding patterns in static images and videos.

▪ Object tracking

▪ Segmentation

▪ Object Detection

▪ Image classification

Page 19: AI and Machine Learning in Healthcare

3 General Applications of AI

Natural Language Processing

▪ Algorithms that process and analyze large pools of human language.

▪ Lexical analysis

▪ Interpret semantics

▪ Pragmatics analysis

Page 20: AI and Machine Learning in Healthcare

3 General Applications of AI

Speech Recognition

▪ Algorithms that interpret and produce sounds.

▪ Voice recognition

▪ Speech-to-text

Page 21: AI and Machine Learning in Healthcare

AI in Diagnostics

A meta-analysis (n=69) published by the Lancet in 2019, showed that health professionals and AI algorithms show a similar sensitivity and specificity for medical diagnostics. https://doi.org/10.1016/S2589-7500(19)30123-2

Page 22: AI and Machine Learning in Healthcare

AI in Diagnostics

Radiology - AI can excel at recognizing patterns in medical images

• Radiology: Artificial Intelligence

Page 23: AI and Machine Learning in Healthcare

AI in Diagnostics

Chatbot - Using natural language processing (NLP), chatbots can acquire patient’s symptoms, triage, and perform a partial diagnosis to aid the healthcare worker.

• Microsoft Healthcare Bot

• Mediktor

Page 24: AI and Machine Learning in Healthcare

AI in Diagnostics

Pathology - Like Radiology, but at cellular level, often incorporating more features such as fluorescent cell markers.

• PathAI

• Paige.AI

Page 25: AI and Machine Learning in Healthcare

AI in Medical Coding

The new ICD-11 has 4x the number of codes compared to the ICD-10 and currently, 51% of human entered codes are inaccurate. AI systems aim to increase both speed and accuracy of the coding procedure.doi: 10.1097/SLA.0000000000000851.

Page 26: AI and Machine Learning in Healthcare

AI in Medical Coding

Interpret multiple data types

• Diagnoss can assess medical images in addition to classifying text.

Page 27: AI and Machine Learning in Healthcare

AI in Medical Coding

On-demand and scalable

• Fathom can regulate itself depending on the workflow, thereby reducing costs.

Page 28: AI and Machine Learning in Healthcare

AI in Medical Coding

Real-Time reporting with dashboards

• 3M™ 360 Encompass™ System has dashboards that produce real-time data for reporting purposes.

Page 29: AI and Machine Learning in Healthcare

AI in Clinical Research

In 2000-2015, just 13.8% of clinical trials passed all three stages of FDA testing. AI aims to speed up the process and increase the success rate.doi: 10.1038/d41586-019-02871-3

Page 30: AI and Machine Learning in Healthcare

AI in Clinical Research

Proper subject attainment• Deep 6 AI uses NLP to process medical records

and reports to find potential subjects in a matter of minutes.

Page 31: AI and Machine Learning in Healthcare

AI in Clinical Research

Direct subject observation

• AICure uses AI in mobile devices to directly monitor patient adherence.

Page 32: AI and Machine Learning in Healthcare

AI in Clinical Research

Design better trials

• trails.ai uses NLP to search previous clinical trials, pertinent regulatory information, and medical journals to design clinical trials.

Page 33: AI and Machine Learning in Healthcare

AI in Clinical Infrastructure

AI can also help with basic operations infrastructure tasks needed by clinical entities.

Page 34: AI and Machine Learning in Healthcare

AI in Clinical Infrastructure

Supply management

• Signant Health- GxP compliant software that digitizes supplies, forecasts quantities, orders supplies, and handles recalls/expired products.

Page 35: AI and Machine Learning in Healthcare

AI in Clinical Infrastructure

Identify patient flow choke points

• Qventus -in real-time can monitor patient flow within a clinic and determine bottlenecks. Additionally, management can create process focus points.

Page 36: AI and Machine Learning in Healthcare

AI in Clinical Infrastructure

Claims and Denials management

• Olive - submits claims faster and more accurately than human, allowing users to concentrate on border cases.

Page 37: AI and Machine Learning in Healthcare

Implementation – Logistical, ethical, and legal considerations to implementing AI.

Page 38: AI and Machine Learning in Healthcare

Implementation - Personnel

Page 39: AI and Machine Learning in Healthcare

Implementation - Cloud vs local inference

Cloud learning

▪ Very powerful and scalable (necessary for onlinemodel)

▪ Can be very expensive▪ 4.6 million to train GPT-3.

▪ End-user must be connected to internet

▪ Possible security concerns

Page 40: AI and Machine Learning in Healthcare

Implementation - Cloud vs local inference

Edge Device

▪ Not very powerful, huge power draw▪ GPT-3 would require 355 years.

▪ Not expensive at all

▪ No internet needed

▪ Cannot create model, only run inference

Page 41: AI and Machine Learning in Healthcare

Implementation – Legal Concerns

HIPAA - One of the most central concerns to multiple facets of integration.

• HIPAA data necessary for developing most medical AI models.

• Needed for inference and consequently, patient data is only as secure as the AI implementation.

Page 42: AI and Machine Learning in Healthcare

Implementation – Legal Concerns

Legal system can’t keep up with the technology

• The FDA is not enamored with online systems.

• Alice Corp. v. CLS Bank International

• Novelty must be found in workflow

Page 43: AI and Machine Learning in Healthcare

Implementation – Legal Concerns

Accountability – What entity is responsible for the AI’s decision?

Page 44: AI and Machine Learning in Healthcare

Implementation – Legal Concerns

Biases – Introduced when creating the model or the result of a model’s inference.

Page 45: AI and Machine Learning in Healthcare

Implementation – Legal Concerns

Transparency – Specifically in deep learning the mechanism of the model is opaque.

Page 46: AI and Machine Learning in Healthcare

Implementation – Ethical Concerns

• Biases – Introduced when creating the model or the result of a model’s inference.

• AI replacing traditional jobs.

• AI used to determine a patient’s medical coverage.

Page 47: AI and Machine Learning in Healthcare

THANK YOU FOR YOUR TIME