Tumori Journal Abstract Book - October 2020 - 10

10
molecular analysis was regularly reimbursed among the
majority (84%) of evaluated centers. KRAS mutations
were detected in 212/766 (27.6%) patients with a newly
diagnosed metastatic lung ADC within the 2019. The
KRAS p.G12C and p.G12V were the most frequent mutations found in 92/766 (12%) and 37/766 (4.8%) of evaluated patients, thus representing about 43% and 17% of the
overall KRAS mutations, respectively. The other mutations found in the tested population included KRAS
p.G12D (2.9%), p.G12A (2.2%), p.G13C (1.2%), p.G13D
(0.9%), p.Q61H (0.9%), p.G12R (0.5%), p.G12S (0.4%),
and p.Q61L (0.3%).
Conclusions: These data revealed that the current status
of KRAS detection in advanced NSCLC patients is quite
heterogeneous in Italy, leading to a different mutation rate
and specific variant distribution. The standardization of
diagnostic molecular approaches across the different
Italian regions represents a major challenges to be adequately addressed.

A11
Machine Learning and prognosis in
lung cancers: a proof of concept
Ponzetti A.1, Zilich R.2, Pisani F.2, Cristiano C.1, Bustreo S.1, Milanesi
E.1, Filippi R.1, Laface R.1, Galliano R.1, Ciuffreda L.1
1
S.C. Oncologia Medica 1, Presidio Molinette, A.O.U. "Città della Salute e
della Scienza", Torino; 2Mix-x S.r.l. - Rulex® Logic Learning Machine, Ivrea

Background: Prognostication is a complex but fundamental part of clinical practice for aggressive tumors of elderly
or frail patients like lung cancers. Unfortunately, in this setting only few independent prognostic factors have been
clearly identified: stage at diagnosis, performance status
(PS) and target-gene alterations. Moreover, especially in
the advanced disease, the relative role of these factors is not
well defined.
Artificial Intelligence (AI)-Machine Learning (ML) techniques, are able to deeply analyze the interactions of multiple clinical variables. This observational, longitudinal,
retrospective study evaluated the ability of ML to identify
the absolute and relative prognostic role of multiple clinical variables in patients affected from lung cancers and
treated in a tertiary care center.
Methods: A clinical database of 592 consecutive patients
diagnosed in 2014-2019 with NSCLC or SCLC was analyzed using a unique ML algorithm: the Rulex 'clear
box' Logic Learning Machine. The absolute and relative
prognostic role of 53 variables was assessed. Then the
rules generated by the system were used to identify the
most relevant prognostic variables together with their
threshold values. Afterwards the system generated a predictive model of 1-year survival from diagnosis,

Tumori Journal 106(2S)
explaining, for each individual, the reason of the prediction (which rule was applied by the model).
Results: Mean age of patients was 67.2 years and 90%
were affected by NSCLC (67.4% adenocarcinoma); about
60% had a stage IV disease at diagnosis,15% a stage IIIA.
Nearly 22% had a PS 2 or higher and 12% had a mutations
of EGFR or an ALK amplification. Globally the prognostic algorithm had an accuracy of 92% with sensitivity and
specificity of 0.88 and 0.95. AUC, NPV and PPV were
respectively 0.92 (IC95% 0.90-0.94), 0.90 and 0.94.
Regarding the identification of prognostic factors, 17
rules have been generated by the system and main prognostic variables were identified. When escluding stage,
PS and molecular alterations, time to treatment initiation
(TTI) and progression-free survival (PFS) had the strongest impact.
In advanced NSCLC, the impact of PS at diagnosis was
strongest than age's impact. A predictive model of 1-year
survival was created and 21 specific "rules" were identified.
Conclusions: When applying AI/ML techiques to prognostication in lung cancers, high accuracy can be reached
with a subsequent possibile strong clinical impact and it
would be worth confirming it with prospective studies.

A12
Liquid biopsy and PET parameters
as predictive factors of
osimertinib treatment in
advanced EGFR-mutated NSCLC
Leonetti A.1, Bola S.2, Minari R.1, Scarlattei M.2, Buti S.1, Bordi P.1,
Baldari G.2, Gnetti L.3, Sammartano A.2, Migliari S.2, Cosenza A.1,
Ferri L.1, Bonatti F.1, Mastrodomenico L.4, Ruffini L.2, Tiseo M.4
1
Medical Oncology Unit, University Hospital of Parma, Parma; 2Nuclear
Medicine Unit, University Hospital of Parma, Parma; 3Pathology Unit,
University Hospital of Parma, Parma; 4Department of Medicine and Surgery,
University of Parma, Parma

Background: Osimertinib resistance constitutes a major
challenge for the management of advanced EGFR-mutated
(adv-EGFR+) NSCLC and the identification of predictive
factors is largely needed. We aimed to investigate a putative role of liquid biopsy (LB) and metabolic parameters in
predicting response to osimertinib.
Methods: We prospectively enrolled adv-EGFR+ NSCLC
patients (pts) who were treated with osimertinib both in
first- (1L) and second-line (2L). Pts underwent LB before
starting of osimertinib and EGFR activating mutation
(EGFRact) allele frequency (AF) was assessed on circulating free DNA (cfDNA) by ddPCR. Pts with EGFRact AF
> 0 were defined as shedders. FDG-PET was performed at
baseline and after 1 month of treatment; metabolic response
(MR) was assessed as per PERCIST 1.0 criteria. Tumor

Tumori Journal Abstract Book - October 2020

Table of Contents for the Digital Edition of Tumori Journal Abstract Book - October 2020