AMS-JSPS-AMED Joint Symposium on Data-Driven Health

AMS-JSPS-AMED Joint Symposium on Data-Driven Health:

Title: Health data applications in clinical care~what can we do using routinely collected data?~

Masao Iwagami, MD, MPH, MSc, PhDDept. of Health Services Research, Univ. of Tsukuba, Japan

Dept. of Non-Communicable Disease Epidemiology, London School of Hygiene and Tropical Medicine, UK

Agenda:1. Introduction (my experience to date)2. What can we do using routinely collected health data? 3. My view on predicting individual risk of an outcome

2008 Univ. of Tokyo (MD)2008-11 Junior and senior residency 2012 Univ. of Tokyo, School of Public Health (MPH)2013 LSHTM, Epidemiology (MSc)2014 LSHTM, Epidemiology and Population Health (PhD)2018-now LSHTM, Honorary Assistant Professor2018-now Univ. of Tsukuba, Assistant Professor

1. Introduction: Masao Iwagami, MD, MPH, MSc, PhD

Inpatient routinely collected data

Outpatient routinely collected data

Key words: Routinely collected health dataAcute kidney injury (AKI), Chronic kidney disease (CKD),Sepsis, Mental health disorders, Pharmacoepidemiology

DPC claims database NDB claims database

1. Introduction: Masao Iwagami, MD, MPH, MSc, PhD

2. What can we do using routinely collected data? (i) To describe burden of a disease(ii) To examine the association between an exposure and

an outcome(iii) To predict individual risk of an outcome

Mortality of dialysis AKI = 50.6%

Pts with CKD are hospitalisedmore often than Pts without CKDfor various reasons

Most likely conclusions: More clinical attention, research, and funding are needed for the disease

(i) To describe burden of a disease(ii) To examine the association between an exposure and

an outcome(iii) To predict individual risk of an outcome

2. What can we do using routinely collected data?

There is association between SSRI (antidepressants) and GI bleeding

There is no association between endotoxin adsorption and mortality

Most likely conclusion: If the association was causal,modifying the exposure would/wouldn’t improve the outcome

(i) To describe burden of a disease(ii) To examine the association between an exposure and

an outcome(iii) To predict individual risk of an outcome

2. What can we do using routinely collected data?

(i) To describe burden of a disease(ii) To examine the association between an exposure and

an outcome(iii) To predict individual risk of an outcome

2. What can we do using routinely collected data?

3. My view on predicting individual risk of an outcome(i) Is it useful? (ii) Is machine learning better than traditional methods?(iii) Does better prediction benefit more?

(i) Is it useful? (ii) Is machine learning better than traditional methods?(iii) Does better prediction benefit more?

3. My view on predicting individual risk of an outcome

Indication for statin

(i) Is it useful? (ii) Is machine learning better than traditional methods?(iii) Does better prediction benefit more?

3. My view on predicting individual risk of an outcome

Risk of incident DM within 3 years

Diabetes prediction tool (Japan)

(i) Is it useful? (ii) Is machine learning better than traditional methods?(iii) Does better prediction benefit more?

3. My view on predicting individual risk of an outcome

(i) Is it useful? (ii) Is machine learning better than traditional methods?(iii) Does better prediction benefit more?

3. My view on predicting individual risk of an outcome

(i) Is it useful? (ii) Is machine learning better than traditional methods?(iii) Does better prediction benefit more?

3. My view on predicting individual risk of an outcome

What is your plan? Your current risk of colon cancer = 70%

Your future risk of stroke = 70%

Bad validitySensitivity = 50%Specificity = 50%

Good validitySensitivity = 95%Specificity = 95%

Keep observation

Resecting colon

Stop smokingExerciseDecrease BP

Stop smokingExerciseDecrease BP

Key messages from Dr. Masao Iwagami

Routinely-collected health data can be used(i) To describe burden of a disease(ii) To examine the association between an exposure and

an outcome(iii) To predict individual risk of an outcome

In the predicting individual risk of an outcome, (i) Is it useful? (ii) Is machine learning better than traditional methods?(iii) Does better prediction benefit more?The most important is to find when the answers are “yes”.