7/22/2014

1

Database sharing model for research and

decision support in radiation therapy

Todd McNutt, Scott Robertson, Sierra Cheng, Joseph Moore, Harry

Quon, Joseph Herman, Michael Bowers, John Wong, Theodore

DeWeese

Disclosure:

Funding from Elekta, Philips and Toshiba

Personalized care using

database of prior patients

Patient

History

Performance

Status

Clinical

assessments

Quality

of Life Survival

Lab Values

Comorbidities

Diagnosis

Disease

Status

Chemotherapy Surgery

Pathology

Toxicities

Radiotherapy

FO

LL

OW

-UP

TREATMENT

CO

NS

UL

T

How to best to

treat individual

patient?

Prediction of

complications for

early

intervention?

Predicted

response and

toxicity?

• Integration of data collection with clinical workflow

– “Big Data” requires meaningful data

• Database design, security and distributed web-access

• Tools for query, analysis, navigation and decision support

– Sample Questions and Uses

• Toxicity Trending

• DVH vs Toxicity

• Automatic Treatment Planning and Quality

• Prophylactic PEG use

Project components

7/22/2014

2

Treatment Timeline

Simulation Targets

OARs

OVH

Consult Demographics

Diagnosis

Staging

Baseline Tox Baseline QoL

History

Planning Rx

Dose

DVH

Weekly On

Treatment Toxicity

QoL

Patient status

Symptom Mgmt

End of

Treatment Acute toxicity

QoL

Patient status

Symptom mgmt

Disease response

Follow Up Late toxicity

QoL

Patient status

Disease response

Image

Guidance Motion

Disease

Response

Chart

Review

Data Collection in Clinic

7/22/2014 5

Clinical Assessment

Quality of life Disease Status

Chart Review (~40 per hour)

7/22/2014 6

Highlight when

data missing

Identify abnormal

prescriptions

7/22/2014

3

Extract, Transform, Load

Oncospace MOSAIQ

Pinnacle TPS

&

DICOM RT PACS

- Scripts, Python, DICOM

- DVH, OVH, Shapes

- SQL Query

- Lab, Toxicity,

Assessments

Oncospace tables and schema

(m:n)

(m:n)

Patient

Family

History

Social

History

Medical

History

Private Health Info

(access restricted)

Tumors

Test Results

(Labs)

Assessments

(Toxicities)

Clinical

Events

Surgical

Procedures

Medications

(chemo)

Organ Dose

Summaries

Radiation

Summary

Patient Representations

(CT based geometries)

Pathology

Feature

Image

Feature

Organ

DVH Data

Organ DVH

Feature

Image

Transform

Radiotherapy

Sessions

Regions of

Interest

ROI Dose

Summary

Shape

Descriptor

Shape

Relationship

Data Features ROI DVH

Features

ROI DVH

Data

1 : N multiple instances 1 : 1 single instance m : n relates m to n

Use of SQL

DB reduces

search to an

SQL query

7/22/2014

4

Designed for data sharing U54 Grant (U Wash, Penn, U Mich)

Head & Neck

700+ pts

Johns Hopkins Institution 1-N (UW)

Panc SBRT

Thoracic

Head & Neck

Panc SBRT

Prostate

shared

Pancreas

150+ pts

Informaticist and the Clinician

• Where clinical knowledge and

informatics science meet?

• What is real knowledge?

• What is NEW knowledge?

7/22/2014 11

The Vs of Big Data

7/22/2014 12

7/22/2014

5

Viability and Value

• Predictive factors must be accessible for new

patients

• Prediction must be clinically valuable and extend

the knowledge of the clinician

7/22/2014 13

Decision support for…

• SAFETY can be improved by alerting users when patient

treatment information deviates from normal.

• QUALITY can be improved by predicting how well you can

do for a patient and seeking to achieve it.

• PERSONALIZATION occurs when physicians and patients

can review results of prior similar patients and make

decisions based on the data specific to the patients needs.

Toxicity trends during and after

treatment – detect outliers

Dysphagia Swallowing

Mucositis Inflammation

Xerostomia Dry Mouth

Worsens after Tx

for many patients

then improves long

term

Heals after Tx for

most patients

Tends to be

permanent

Follow up During Treatment

#pts

Toxicity grade (0-5)

7/22/2014

6

Shape-dose relationship for

auto-planning

Decisions:

Plan quality assessment

Automated planning

Expected toxicities

Dosimetric trade-offs

• More efficient plan optimization (10 fold)

• Normal tissue doses reduced (5-10%)

• Clinically released for Pancreatic Cancer

Shape relationship Dose prediction DB of prior patients

parotids

PTV 0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Dose (Gy)

no

rmaliz

ed v

olu

me

Right parotid

Left parotid

Flavors of Data Depends of timeline

FIXED

already

happened

VARIABLE

influence

outcome

measures

POPULATION

predict

outcome

measures

Use to stratify patients Adjustable to

influence outcome?

Predict outcome with

FIXED and VARIABLE

based on POPULATION?

Treatment Timeline

Simulation Targets

OARs

OVH

Consult Demographics

Diagnosis

Staging

Baseline Tox Baseline QoL

History

Planning Rx

Dose

DVH

Weekly On

Treatment Toxicity

QoL

Patient status

Symptom Mgmt

End of

Treatment Acute toxicity

QoL

Patient status

Symptom mgmt

Disease response

Follow Up Late toxicity

QoL

Patient status

Disease response

Image

Guidance Motion

Disease

Response

Auto

Plan

Risk

Based

Symptom

Mgmt

Therapy

Mgmt

At what time point do we have

enough data to make decision

based on future prediction?

Input Variables => Prediction?

7/22/2014

7

DVH, Toxicities and Grade

distributions

Voice Change

Larynx

50% Volume

Dysphagia

Larynx_edema

30% Volume

Number of

patients by

grade at D50%

Toxicity Grade

0,1,2,3,4,5

Mean and stddev

of DX% at grade

Dysphagia and Xerostomia

Probability of Grade 2 Dysphagia

Dose (cGy)

No

rma

lize

d V

olu

me

0 1000 2000 3000 4000 5000 6000 70000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 1000 2000 3000 4000 5000 6000 70000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

All DVH Curves for the Larynx (142 Patients)

Dose (cGy)

No

rma

lize

d V

olu

me

Dysphagia Grade < 2 (N=87)

Dysphagia Grade 2 (N=55)

Larynx vs Grade ≥ 2 Dysphagia

Plugging Into Oncospace Scott Robertson PhD

21

7/22/2014

8

Voice Change

Bad DVH!

• DVH assumes that every sub-region of an OAR has the

same radiosensitivity and functional importance to the

related toxicity

• DVH assumes that each OAR is uniquely responsible for

the overall human function related to the toxicity

≠

Parotid Data

7/22/2014

9

Classification with correlated features: unreliability of

feature ranking and solutions

http://www.cedars-sinai.edu/

Simulation of 1, 10 and 20 variables with a correlation of

0.9 with variable 3

Genuer et al.

Correlation is not Causation

7/22/2014 26

Acknowledgments

• JHU - CS

– Russ Taylor PhD

– Misha Kazhdan PhD

– Patricio Simari PhD

– Jonathan Katzman

• JHU - Physics

– Alex Szalay PhD

– Tomas’ Bodavari PhD

• Philips PROS

– Karl Bzdusek

• Erasmus

– Steven Petit PhD

• JHU-RO

– Binbin Wu PhD

– Kim Evans MS

– Robert Jacques PhD

– Joseph Moore PhD

– Scott Robertson PhD

– Wuyang Yang MS, MD

– John Wong PhD

– Theodore DeWeese MD

– GI Team

– Joseph Herman MD

– Amy Hacker-Prietz PA

– H&N Team

– Harry Quon MD

– Giuseppe Sanguineti MD

– Heather Starmer MD

– Jeremy Richmond MD

– Anna Keiss MD