What is the real value of predictive models?

Identification of image-based phenotypes that can be combined with clinical data and emerging biomarkers has the potential for great clinical impact. The development of these predictive models will not only improve evaluation, risk stratification, diagnosis and treatment of non-small cell lung cancer (NSCLC) patients, but also, make clinicians work easier, allowing for an integrated and practical analysis of all patients’ characteristics (1).

However, regardless of the increasing amount of data obtained from each patient, clinical utility and implementation of these models remain controversial and more challenging than initially expected.

Hannequin et al. (2) present the results of their retrospective study where positron emission tomography/computed tomography (PET/CT) and CT scan radiomic features fail to predict overall survival (OS) and progression-free survival (PFS) when compared to other traditionally used clinical variables, as gender and stage.

In this issue of Translational Lung Cancer Research, Hannequin et al. describe the results of different predictive models including clinical characteristics, as gender and tumor stage, radiomic features obtained from CT scan and PET-CT images and histological features as PD-L1 expression. Despite the intrinsic power author’s observe in CT and PET-CT radiomic features to predict survival, the truth is their models are not able to predict OS and PFS better than traditional models based in clinical data.

As the authors describe, among the main limitations of their study are its retrospective nature, the low number of events included in their stepwise models, that could have affected the results, the low number of patients studied and the lack of consideration of other important histopathological characteristics, as response to immune checkpoint inhibitors (ICI) or neutrophils to leukocyte ratios.

Furthermore, although predictive models trying to correlate radiomic features to clinical and histopathological responses to treatment continue to proliferate, the role of radiomics in real clinical practice remains still indeterminate.

Predictive models based in clinical, radiologic and genomic features, will hopefully improve patients’ data processing and help in multidisciplinary teams’ decision making, taking into account specific patients’ characteristics (1).

Unfortunately, despite multiple studies trying to correlate radiological findings with tumor characteristics, OS and PFS, they still face some challenges and potentially difficult to overcome limitations, as radiomic workflow lack of repeatability or reproducibility under particular circumstances (3). In addition, although the field of radiogenomics aims to combine radiomic features with genetic biomarkers (4), the major challenge it faces is the difficulty to obtain comprehensive information regarding tumor biology and responses to treatment. Whether this challenge could be overcome using artificial intelligence to analyze large imaging and pathological datasets would be determined by costs, time and data analysis limitations.

In the present study, authors have tried to correlate CT and PET-CT findings with PD-L1 expression, and, although other investigators have tried to find a relationship between CT and PET-CT characteristics with EGFR, KRAS mutations and ALK translocations (5-7), their results remain difficult to translate into clinical practice. Other attempts to try to predict OS and PFS in early stage NSCLC, as the model designed and validated by Vaidya et al. (8) have shown promising results, even in predicting the benefit of adjuvant chemotherapy following surgical resection.

The number of predictive models developed in the recent years focused in both improvement of diagnosis and prediction of survival in NSCLC, has increased exponentially, but their clinical applications and their usefulness in a real word scenario is yet to be defined. Similarly, the number of variables and features included in these models does not correlate to their ability to discriminate different outcomes, and the simpler the model, the more likely it would be implemented in daily clinical practice.

To overcome the current challenges, it is of utmost importance to have prospective, independent and properly audited databases able to capture clinical, radiologic, surgical and genetic information in a standardized, reproducible manner. Furthermore, the problem that entails the processing of the massive amount of data generated will be solved, or at least improved, by the correct application of these models (1,9).

In a near future where machine-learning processes will be part of the healthcare workforce, the correct development and integration of these predictive models in the daily clinical practice will not only facilitate all clinicians’ work but also improve considerably patient care, from diagnosis to treatment and follow-up, avoiding unneeded invasive procedures and derived complications.

Acknowledgments

Funding: None.

Footnote

Provenance and Peer Review: This article was commissioned by the editorial office, Translational Lung Cancer Research. The article did not undergo external peer review.

Conflicts of Interest: Both authors have completed the ICMJE uniform disclosure form (available at https://tlcr.amegroups.com/article/view/10.21037/tlcr-22-735/coif). MR reports fees not related to this manuscript from Abex and AstraZeneca. LM has no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Ather S, Kadir T, Gleeson F. Artificial intelligence and radiomics in pulmonary nodule management: current status and future applications. Clin Radiol 2020;75:13-9. [Crossref] [PubMed]
2. Hannequin P, Decroisette C, Kermanach P, et al. FDG PET and CT radiomics in diagnosis and prognosis of non-small-cell lung cancer. Transl Lung Cancer Res 2022;11:2051-63. [Crossref] [PubMed]
3. Fornacon-Wood I, Faivre-Finn C, O'Connor JPB, et al. Radiomics as a personalized medicine tool in lung cancer: separating the hope from the hype. Lung Cancer 2020;146:197-208. [Crossref] [PubMed]
4. Lo Gullo R, Daimiel I, Morris EA, et al. Combining molecular and imaging metrics in cancer: radiogenomics. Insights Imaging 2020;11:1. [Crossref] [PubMed]
5. Glynn C, Zakowski MF, Ginsberg MS. Are there imaging characteristics associated with epidermal growth factor receptor and KRAS mutations in patients with adenocarcinoma of the lung with bronchioloalveolar features? J Thorac Oncol 2010;5:344-8. [Crossref] [PubMed]
6. Lee Y, Lee HJ, Kim YT, et al. Imaging characteristics of stage I non-small cell lung cancer on CT and FDGPET: relationship with epidermal growth factor receptor protein expression status and survival. Korean J Radiol 2013;14:375-83. [Crossref] [PubMed]
7. Rizzo S, Petrella F, Buscarino V, et al. CT radiogenomic characterization of EGFR, K-RAS, and ALK mutations in non-small cell lung cancer. Eur Radiol 2016;26:32-42. [Crossref] [PubMed]
8. Vaidya P, Bera K, Gupta A, et al. CT derived radiomic score for predicting the added benefit of adjuvant chemotherapy following surgery in Stage I, II resectable Non-Small Cell Lung Cancer: a retrospective multi-cohort study for outcome prediction. Lancet Digit Health 2020;2:e116-28. [Crossref] [PubMed]
9. Jha S, Topol EJ. Adapting to Artificial Intelligence: Radiologists and Pathologists as Information Specialists. JAMA 2016;316:2353-4. [Crossref] [PubMed]