Digital Health Innovation and Informatics

PD 14 - Digital Health Information & Informatics - Poster Discussion

1121 - Learning a Cox Model Predicting Survival Based on 3413 Routine Clinical Rectal Cancer Patients Without Sharing Patient Data

Tuesday, October 23
5:09 PM - 5:15 PM
Location: Room 217 A/B

Learning a Cox Model Predicting Survival Based on 3413 Routine Clinical Rectal Cancer Patients Without Sharing Patient Data
A. Damiani1, C. Masciocchi1, J. van Soest2, N. Dinapoli3, J. Lenkowicz1, G. Chiloiro1, M. A. Gambacorta4, B. Corvari3, E. Meldolesi3, A. R. Alitto3, L. Boldrini1, A. Dekker2, and V. Valentini1; 1Polo Scienze Oncologiche ed Ematologiche, Istituto di radiologia, Università Cattolica del Sacro Cuore-Fondazione Policlinico Universitario A.Gemelli, Roma, Italy, 2GROW School for Oncology and Developmental Biology, Maastricht University Medical Centre, Maastricht, Netherlands, 3Polo Scienze Oncologiche ed Ematologiche, Università Cattolica del Sacro Cuore - Fondazione Policlinico Universitario Agostino Gemelli, Roma, Italy, 4Polo Scienze Oncologiche ed Ematologiche, Istituto di Radiologia, Università Cattolica del Sacro Cuore-Fondazione Policlinico Universitario Agostino Gemelli, Roma, Italy

Purpose/Objective(s): Developing a prediction model (PM) to support personalized medicine is a challenge on many different levels. Multi-institutional studies, a suitable solution, are often hampered by ethical, political and legal problem involving data sharing. In this work, we applied the distributed learning approach to overcome these issues. The aim of this work is to train a PM using a large number of rectal cancer (RC) patient data coming from 2 different European centers without data leaving the hospital, including the option to regularly re-learn the PM.

Materials/Methods: Clinical data of RC patients (pts) were regularly extracted from source systems (e.g. EMR and TPS), de-identified and translated into a semantically interoperable format using Semantic Web technologies. An online learning portal (OLP) managed the message delivery between sites, where a written custom distributed version of the Cox Proportional Hazard (PH) regression model was iteratively fitted to data. The clinical tumoral (cT), nodal stage (cN), gender and age at diagnosis were considered as predictive factors in an overall survival (OS) model. The hazard ratio (HR) and the p-value for each predictor were evaluated. A conservative p-value 0.01 was considered significant for variable selection, due to the high number of available pts. The distributed version of the Cox PH, was tested before to prove mathematically identical to its traditional, centralized counterpart.

Results: A real distributed PM was trained on 3413 RC from 2 different institutions (see table 1) without sharing clinical data. cN was the main predictive factor of the PM (HR=1.03 and p-value < 0.01). cT (HR=1.2 and p-value < 0.01) and age (HR=1.4 and p-value < 0.01) also affect the prediction of OS. Gender was found not statistically significant (p=0.04). Performance of the IT infrastructure was adequate to the goal, and the semantic approach to remapping pre-existing terminological systems into a shared data representation using Semantic Web technologies proves feasible and effective.

Table 1. Patient characteristics
Covariate distribution

Pts number

Institution 1

Institution 2

3413

617

2796

Gender

Female

Male

_

36%

64%

cT

1-2

3

4

_

21%

67%

12%

cN

0

1

2

3

_

27%

35%

36%

2%

OS

Death

Alive

_

33%

67%

Median OS time [months]

70

Conclusion: To the best of our knowledge, this is the first Cox PH model retrospectively trained on a large set of RC pts from different institutions. A distributed learning approach was used based on an industry-level, freely available IT infrastructure (OLP), including a semantic approach to achieve a uniform, computer-interpretable data representation. This analysis can be repeated regularly, as the hospital’s data may increase over time.

Author Disclosure: A. Damiani: None. C. Masciocchi: None. J. van Soest: None. N. Dinapoli: None. A. Dekker: Research Grant; Varian Medical Systems. Honoraria; Varian Medical Systems. Consultant; Varian Medical Systems. Chief Scientific Officer; Medical Data Works B.V. Member scientific advisory board; Peter Munk Cardiac Centre.

Carlotta Masciocchi, MSc

Disclosure:
No relationships to disclose.

Presentation(s):

Send Email for Carlotta Masciocchi


Assets

1121 - Learning a Cox Model Predicting Survival Based on 3413 Routine Clinical Rectal Cancer Patients Without Sharing Patient Data



Attendees who have favorited this

Please enter your access key

The asset you are trying to access is locked. Please enter your access key to unlock.

Send Email for Learning a Cox Model Predicting Survival Based on 3413 Routine Clinical Rectal Cancer Patients Without Sharing Patient Data