Click here to


Are you sure ?

Yes, do it No, cancel

Non-Small Lung Cancer Survival Decision Support System Model Trained Using Distributed Learning Over a Multi-Institutional Cohort

D Thwaites1*, M Field2 , L Holloway3 , J Lehmann4 , J Sykes5, (1) Institute of Medical Physics, The University of Sydney, Sydney, NSW, (2) University of New South Wales,Sydney,NSW (3) Liverpool and Macarthur cancer therapy centres and ingham Institute, Sydney, NSW, (4) Newcastle Mater Hospital, Hunter Region Mail C, (5) Blacktown Cancer and Haematology Centre, Sydney, NSW,


(Sunday, 7/29/2018) 2:05 PM - 3:00 PM

Room: Karl Dean Ballroom A1

Purpose: Clinical decision support systems (DSS) combine multiple variables in a statistical analysis to predict treatment outcomes. The selection of radiotherapy (RT) treatment of non-small cell lung cancer varies in practice, with different proportions of curative versus palliative RT use. We updated an overall survival DSS model for patients treated with curative RT by including data from additional institutes in the analysis. Data was stored in three institutions and the model was built using a distributed learning algorithm.

Methods: Clinical information and RT planning computed tomography (CT) data for 810 patients with inoperable, Stage I-III NSCLC treated with RT between 2003 and 2017 were compiled at three institutions. There were 511 patients at Institute 1, 96 at Institute 2 and 203 at Institute 3. Of these patients 466 received curative RT (dose > 48Gy). A support vector machine for predicting two-year survival was trained across the three institutions using distributed learning software on a randomly-selected half of the cohort. Assignment to one of three risk groups was adjusted towards providing treatment selection decision support between curative and palliative treatment based on the training cohort. The model was tested on the remaining half of the cohort. The attributes used were age, gender, performance status, tumour volume.

Results: The model had a 0.65 AUC on the withheld data. Analysis for risk groups is displayed in Figure 1, including the patients who received palliative treatment. Here, 16% (n=27) of patients treated with palliative RT resided in the predicted low risk group and observed significantly reduced survival (logrank, p<0.01) over curative treatment. Alternatively, 12% (n=29) of curatively treated patients belonged to the predicted high-risk group and did not observe increased survival with curative RT (logrank, p=0.075).

Conclusion: The distributed model exhibited competitive performance and is a novel approach to building models informing treatment selection.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email