From April 25-27th 2019 Hospital Israelita Albert Einstein (HIAE), the PROADI-SUS project and MIT organized its Artificial Intelligence in Healthcare Symposium and Datathon.

In the first day, we attended a Symposium including talks and round tables with experts in the field of Artificial Intelligence, Big Data and Healthcare. I took some notes of things I found interesting and will be posting here for further consultation.

Symposium

Richard Delaney - Vital Strategies

Intersection Between High-dimensionality Data and Established Public Health Databases

High Dimensional Data:
- Many factors to control for, harder to do hypothesis testing. One alternative is to leverage machine learning to do more pattern recognition
- Don’t always come with pre conceived ideas, but try to find patterns freely
- Find probabilities of actions to know where to deliver Public health service
Creating conditions to success:
1. Start with a project: avoid “We prepare and then we act”, we are never prepared enough! Start easy, purposeful, datathon events follow this idea!
2. After the first step, discuss systemic changes

Nick Guldemond - The Institute of Health Policy & Management Erasmus University

Digital Technologies for Public and Private Primary Care

Design Thinking for Health Data Science Projects
- Patient Journey
- Clinical Pathway
Data new vs old Wikipedia
- Findable
- Accessible
- Interoperable
- Reusable
If we find a good solution, how do we scale? To another region, etc…

Jacson Venâncio Barros - President of the Brazilian Association of CIO’s in Health

Technologies for Supporting Public Health Management and Governance Problem: Startups are disconnected from end users

70% of hospitals don’t have eletronic medical records
Lack of standards in vocabulary (taxonomy)
Asymmetry of information: each specialty has its own jargon
Lack of trust among health sectors Challenge: Identify in this sea of data, which is the drop that interests us

Round Table: Challenges of Big Data and AI in Population Health - Gisele Bohn, Richard Delaney, Nick Guldemond, Jacson Barros, Fatima Marinho

Topics:

Aggregate behavioral data to clinical data
Think small first: break problems into small questions that big data can help
Predict events that are repeated in certain population groups
Integration with the real world
How social networks influence behaviour

Open Session: Data Access, Privacy and Security - Andrea Suman, Silvio Pereira, Anderson Soares, Marcelo Felix, Rogéria Leoni Cruz

Restrictive law?
Big Data x Data Governance
No safe culture, territoriality, data sharing (would it restrict the access for research?)
Loss of potential of using machine learning

Leonardo Rolim Ferraz

Data Driven ICU - Einstein Initiative

Lucas Bulgarelli - MIT/HIAE

Big Data 360

MIMIC
De-identification
Leveraging electronic health records for clinical research
Sharing - Physionet

Ary Serpa Neto - HIAE

Innovative Models of Clinical Trials Using Large Databases

Christopher Cosgriff - MIT

Deep Learning: A Brief Overview for Clinicians

Round Table: Challenges of Big Data and AI in the Intensive Care Unit - Leo Anthony Celi, Alistair Johnson, Leonardo Rolim Ferraz, Ary Serpa Neto, Lucas Bulgarelli

The Book of Why

Leo Anthony Celi - MIT

MIT Experience in Data Analytics applied to Intensive Care Units

MIMIC (Medical Information Mart for Intensive Care)
Book: Secondary Analysis of Electronic Health Records
Opportunities for AI in Healthcare
- Classification
- Prediction
- Optmization (Precision Medicine)

Alistair Johnson - MIT

Keynote: MIMIC Across Modalities: X-rays and Beyond

Chest X-rays are ubiquitous, radiologists are not
Perfect algorithms on imperfect reports

Matthieu Komorowski - Imperial College London

Keynote: Reinforcement Learning Approaches to Decision Support in Sepsis

Open Session: Opportunities for Innovation in Healthcare - Pedro Marton Pereira, Gustavo Landsberg, Gisele Bohn, André Bem

I forgot to take notes.

Datathon

group9

Cauê Bueno, Eduardo Casaroto, Fernando Ramos, Gustavo Silveira, Jacqueline Silva,
Matheus Silva, Marcelo Fiorelli, Sonia Altavila, Wellington Araújo

Pulse Pressure: a new outcome predictor in the Intensive Care Unit?

Brief literature review and motivation for study

Mean Arterial Pressure (MAP) is the main parameter used to define hemodynamic condition in critically ill patients and levels below 65mmHg are related to poor outcomes especially in septic patients
Pulse Pressure (PP) is related to pressure and stroke volume.
Pulse Pressure is a neglected hemodynamic parameter at bedside.

Aim of Study

Evaluate if Pulse Pressure is a reliable predictor of 28-days mortality compared to Mean Arterial Pressure in critically ill patients

Data Source

Inclusion criteria:
- First ICU admissions (MIMIC-3)
Exclusion criteria
- Patients readmitted at ICU in the same hospitalization
- Age < 18 years

Statistical analysis

Description of the sample:
- mean and standard deviation
- bars plots and boxplots
Logistic regression model
- Outcome: 28-days mortality (from ICU admission)
- Model 1: mean arterial pressure
- Model 2: pulse pressure

Results

Characteristis of the Sample

chartable

SD: standard deviation *Values taken when the minimum systolic blood pressure where observed in the first 24h of ICU admission

MAP vs PP

rplots

pyplots

Logistic Regression

Outcome = 28-days mortality

Model 1: predictor = mean arterial pressure OR = 0.9694 95% IC = 0.9672 - 0.9717

Model 2: predictor = pulse pressure OR = 0.9960 95% IC = 0.9942 - 0.9977

Discussion and Next Steps

MAP assessment alone might not be useful to predict outcome to all ICU patients
Preliminary result: PP assessment looks to be a better outcome predictor for “higher PP” patients
Promising results: min PP>70 - No min MAP difference between outcomes

disc

Next Steps:

Group analysis: who are those patients? Are they elderly people?
Check ICU interventions (mechanical ventilation, vasopressors etc)?
Should we develop a specific guideline for those patients?