1 TM PIONEERING BIG DATA IN RADIATION ONCOLOGY Todd McNutt PhD Associate Professor Radiation Oncology Johns Hopkins University Presented at the Target Insight Toronto, CAN May 8 th, 2015
2 Disclosures This work has been partially funded with collaborations from: Philips Radiation Oncology Systems Elekta Oncology Systems Toshiba Medical Systems as well as Commonwealth Foundation Maritz Foundation 2
3 Problem Ability to advance radiotherapy is limited by our knowledge of which patients are at risk of high grade toxicity or of limited ability to cure. Knowledge from clinical trials is quite coarse and fails to consider all of the aspects of the individual patient. Big Data offers an opportunity to better predict treatment outcome and provide improved clinical decisions for individual patients.
4 Personalized care using database of prior patients How to best to treat individual patient? Pathology Diagnosis Performance Status Comorbidities Patient History Radiotherapy Surgery Chemotherapy CONSULT TREATMENT Survival FOLLOW-UP Quality of Life Clinical assessments Lab Values Toxicities Disease Status Prediction of complications for early intervention? Predicted response and toxicity?
5 Precision Radiotherapy Treatment Planning
6 Mucositis data collected at JHU # patients Toxicity grade 6
7 Promote Culture of Data Collection Data collected over entire treatment Consult Demographics Diagnosis Staging Baseline Tox Baseline QoL History 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 Simulation Targets OARs OVH Planning Rx Dose DVH Image Guidance Motion Disease Response At what time point do we have enough data to make decision based on future prediction? Auto Plan Risk Based Symptom Mgmt Therapy Mgmt Input Variables => Prediction?
8 MOSAIQ Browser Departmental Analytics Framework Multi-Disciplinary Clinics RadOnc Users Security Certificate Johns Hopkins Enterprise Login SiteMinder Access Management Mobile Solutions Mosaiq Browser -IIS -ASP.NET -C# -SQL Data Access Layer SQL SQL Query Log Clinical Analytics IDE - Visual Studio - SVN repository MOSAIQ Security Rights MOSAIQ DB
9 Data Collection in Clinic Clinical Assessment Quality of life Disease Status 5/27/2015 9
10 Head and Neck Oncospace Landscape Home-based Assessment (FitNinja) Head and Neck Tissue Bank (Chung) SNP Genotyping Pre-Treatment Imaging (Kiess/Hobbs/Shen) On-Treatment Imaging (Lee/Shekhar) (Toshiba) Oncospace - On-Treatment Evaluation - RT dosimetry Swallow Function (Starmer et al) Toxicity PROs Oncologic Outcome Neck Function (Sweet) Ocular Function (Subramanian) Post-Treatment Oncologic and Toxicity (Ding/Fakhry/Murano/Prince/ Wasserman Ecological Momentary Assessments (EMAs)
11 Extract, Transform, Load DICO M MOSAIQ - SQL Query - Lab, Toxicity, Assessments Oncospace Pinnacle TPS - Scripts, Python, DICOM - DVH, OVH, Shapes
12 Oncospace tables and schema Patient Private Health Info (access restricted) Family History Social History Medical History Medications (chemo) Surgical Procedures Test Results (Labs) Assessments (Toxicities) Clinical Events Tumors Radiation Summary Patient Representations (CT based geometries) (m:n) Image Transform Image Feature Pathology Feature Organ DVH Data Organ Dose Summaries 1 : N multiple instances 1 : 1 single instance m : n relates m to n Organ DVH Feature ROI DVH Data Radiotherapy Sessions ROI Dose Summary ROI DVH Features Shape Descriptor Regions of Interest Data (m:n) Shape Relationship Features
13 Data inventory Toxicities (24var/pt) Quality of Life (64var/pt) Measured data (6 var/pt) Disease response(8var/pt) w 2-4w 4-6w 6w-3m 3m-6m 6m-1y 1-2y 2-3y 3-4y 4-5y 5-6y >6y 5/27/
14 Organs at risk with full dose 5/27/
15 Learning health system can t exist without data Decision Point Facts Controls Outcomes time Knowledge Database Facts Decisions Controls Presentation of Predictions Predictive Modeling Predicted Outcomes Data Feedback (Facts, Outcomes) 15
16 Oncospace Consortium Repository (It s all about the data) Johns Hopkins U. Washington U. Toronto Sunnybrook U. Virginia Institution X Knowledge Base Registry $/pt Quality Reporting Decision Support Research N
17 Toxicity trends during and after treatment detect outliers During Treatment Follow up Dysphagia Swallowing Worsens after Tx for many patients then improves long term Mucositis Inflammation Heals after Tx for most patients Xerostomia Dry Mouth Tends to be permanent #pts Toxicity grade (0-5)
18 Toxicity trends during and after treatment detect outliers During Treatment Follow up Dysgeusia Taste disturbance Worsens through Tx then improves long term Voice Change Worsens through Tx #pts Toxicity grade (0-5)
19 DVH, Toxicities and Grade distributions Voice Change Larynx 50% Volume Dysphagia Larynx_edema 30% Volume Toxicity Grade 0,1,2,3,4,5 Mean and stddev of DX% at grade Number of patients by grade at D50%
20 OVH: serial vs parallel Target OAR1 OAR2 r 70,1 r 70,2 For parallel organs, OAR2 is more easily spared. For serial organs, OAR1 is more easily spared.
21 Shape-dose relationship for radiation plan quality Shape relationship DB of prior patients Dose prediction parotids 0 PTV normalized volume Dose (Gy) Right parotid Left parotid For a selected Organ at Risk and %V, find the lowest dose achieved from all patients whose %V is closer to the selected target volume? Decisions: Plan quality assessment Automated planning IMRT objective selection Dosimetric trade-offs
22 Sample automated radiation planning result Original plan Automated plan 30% reduction in dose to parotids Auto plan Original Plan Dot: right No-dot: left Dose(Gy) brain (Gy) (max) Brainstem (Gy) (max) Cord4mm (Gy) (max) L inner ear (Gy)(mean) original re-plan R inner ear (Gy) (mean) mandible (Gy) (max) larynx for edema (V50) esophagus (Gy)(max) original % re-plan % 61
23 Toxicity and Dose Volume Histogram Scott Robertson PhD 23
24 Prevalence and prediction of treatment-related complications RISK STRUCTURES TOXICITIES / OUTCOMES 10%/Gy 5%/Gy MODELLED RISK May 27,
25 Voice Change
26 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
27 Spatial dose analysis Method Voice dysfunction n=99, n + =8, n - =91 Xerostomia n=364, n + =275, n - =89 Bagged Naïve Bayes (1000 iterations) Bagged Linear Regression (1000 iterations) Naïve Bayes Linear Regression Random Forest (1000 trees) Hopkins NTCP inhealth LKB
28 Classification with correlated features: unreliability of feature ranking and solutions Simulation of 1, 10 and 20 variables with a correlation of 0.9 with variable 3 Genuer et al.
29 Weight loss prediction Endpoint: > 5kg loss at 3 months post RT Larynx salivary glands thyroid hypopharynx yes FALSE FALSE At planning Larynx D55 < 27.5Gy no Diagnostic ICD-9 Combo Parotid D50< 13.5Gy Distance: HD-PTV to Superior Constrictor Muscle >= 1.1cm Combo Masticatory Muscle D41 < 38.4Gy FALSE FALSE FALSE FALSE oropharynx tongue nasopharynx nasal cavities Larynx D80 < 23.0Gy TRUE Larynx salivary glands thyroid hypopharynx Nausea < 1 (CACAE) FALSE Skin Acute < 3 (CACAE) N stage < 2 Diagnostic ICD-9 At end of RT yes Combo Parotid D61< 7.5Gy Pain Intensity - Current < 5 Able to eat foods I like >= 3 (FACT) larynx D59 < 27.4Gy Distance: LD-PTV to Larynx >= -1.3cm Cricopharyngeal Muscle D100 < 38.1Gy oropharynx tongue nasopharynx nasal cavities Diagnostic ICD-9 tongue Esophagitis/ Pharyngitis < 3 (CACAE) Combo Parotid D96< 6.5Gy Larynx D10 < 42.3Gy FALSE TRUE TRUE FALSE FALSE TRUE FALSE TRUE FALSE TRUE FALSE TRUE FALSE TRUE no 5/27/
30 What can we do with the data? Shape based auto-planning Clinical (prostate, pancreas) Efficient high quality plan Weight loss prediction Improved symptom management Toxicity Risk DVH based Spatial dose based Disease response prediction Pancreas resectability Head and neck HPV dose de-escalation Shape relationship parotids PTV DB of prior patients normalized volume Dose prediction Right parotid Left parotid Dose (Gy) 30
31 Summary The Oncospace model can house RT data effectively and provides a model for sharing Data collection in the clinical environment has been demonstrated All patient on trial Data exploration and analysis across multiple institutions is possible Exploring models to stratify patients to improve the predictive power of the data Decision support to improve quality and safety has been demonstrated Personalized medicine has not been fully demonstrated, but remains a tenable goal
32 Acknowledgments JHU-RO Sierra Cheng MD Kim Evans MS Michael Bowers BS 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 Ana Keiss MD Mysha Allen RN Sara Afonso RN JHU - CS Russ Taylor PhD Misha Kazhdan PhD Fumbeya Murango BS Philips PROS Karl Bzdusek BS Toshiba Minoru Nakatsugawa PhD Bobby Davey PhD Rachel-Louise Koktava John Haller Elekta