Thèse Développement et Validation d'Algorithmes Permettant d'Identifier les Séquelles à Long Terme chez les Survivants d'Un Cancer de l'Enfant à Partir du Chaînage Fccss-Snds Application à l'Étude H/F Doctorat.Gouv.Fr

Établissement : Université Paris-Saclay GS Santé publique École doctorale : Santé Publique Laboratoire de recherche : Centre de Recherche en épidémiologie et Santé des populations Direction de la thèse : Rodrigue ALLODJI ORCID 0000000218958415 Début de la thèse : 2027-10-01 Date limite de candidature : 2026-11-25T23:59:59 Les progrès majeurs réalisés dans le traitement des cancers pédiatriques ont permis d'atteindre des taux de survie à 5 ans supérieurs à 80 %. Cependant, cette amélioration s'accompagne d'une augmentation des séquelles à long terme, incluant notamment les seconds cancers et les pathologies cardiovasculaires, constituant un enjeu majeur de santé publique. Par ailleurs, l'émergence récente de thérapies innovantes (thérapies ciblées, immunothérapies, thérapies cellulaires) soulève de nouvelles questions quant à leurs effets iatrogènes à long terme, encore largement méconnus, en particulier chez les enfants et les adolescents.
Dans ce contexte, cette thèse vise à développer et valider des algorithmes d'identification des séquelles tardives à partir des données du Système National des Données de Santé (SNDS), puis à les appliquer à l'étude des effets des traitements innovants en oncologie pédiatrique.
Le premier axe consistera à construire et valider ces algorithmes en s'appuyant sur la cohorte FCCSS qui inclue 7 670 patients diagnostiqués pour un cancer solide ou un lymphome, disposant de données cliniques validées servant de référence. Des méthodes avancées d'analyse de trajectoires de soins (analyse de séquences, modèles de Markov, apprentissage automatique) seront utilisées pour modéliser les parcours de soins et améliorer l'identification des événements de santé. Les performances des algorithmes seront évaluées à l'aide d'indicateurs de classification.
Le deuxième axe portera sur l'étude des effets iatrogènes des thérapies innovantes dans la cohorte HORUS, qui regroupe des données cliniques détaillées issues de plusieurs centres de lutte contre le cancer, couplées au SNDS. Des analyses de survie (modèles de Cox, modèles à risques concurrents) permettront d'estimer les risques associés à ces expositions, en comparaison aux traitements conventionnels. En complément des méthodes basées sur les scores de propension (IPTW, appariement), une approche d'émulation d'essai cible (target trial emulation) sera considérée.
Enfin, le troisième axe visera à généraliser les résultats à l'échelle nationale en utilisant les données exhaustives du SNDS, afin d'estimer la prévalence du recours thérapies innovantes et des séquelles dans la population française, tout en tenant compte des biais de sélection.
Ce projet s'inscrit à l'interface de l'épidémiologie, de la science des données et de l'oncologie, et contribuera à une meilleure compréhension des effets à long terme des traitements anticancéreux notamment les effets iatrogènes des thérapies innovantes (thérapies ciblées, immunothérapies, thérapies cellulaires), dans une perspective d'amélioration de la prise en charge et du suivi des survivants de cancer de l'enfant, de l'adolescent et du jeune adulte. Over the past five decades, major advances have been made in the treatment of paediatric cancers, leading to a dramatic improvement in prognosis. Overall 5-year survival rates now exceed 80% in children and adolescents. In Europe, there are estimated to be more than 23 million cancer survivors, of whom nearly 600,000 are survivors of paediatric cancers [1]. This improvement in survival nevertheless brings new public health challenges. The long-term care of paediatric cancer survivors is now a priority, included in particular in the French Cancer Plan since 2014 and in the 2021-2030 Ten-Year Cancer Control Strategy (Action IV.2.11). These patients were exposed to treatments that were often intensive and multimodal, resulting in substantial late morbidity: about 68% of adults treated for cancer in childhood have at least one severe or disabling chronic condition [2], and their mortality remains 8 to 11 times higher than that of the general population [3]. Among the most frequent and serious late complications are second cancers, as well as cardiovascular and cerebrovascular diseases [4-6]. These findings rely in part on the work of the Radiation Epidemiology team (EpiRad) of the INSERM UMR1018 unit (CESP), internationally recognised for its research on the late effects of anticancer treatments, notably based on the FCCSS cohort (French Childhood Cancer Survivorship Study), enriched by its linkage with data from the French National Health Data System (SNDS).
Since the 2000s, therapeutic strategies in oncology have changed profoundly with the emergence of innovative treatments, including targeted therapies, immunotherapies and cell therapies (such as CAR-T cells). Initially reserved for early-phase trials in patients with treatment failure, these approaches are now part of standard care, including first-line treatment, and including in paediatric oncology. However, their long-term effects remain largely unknown, particularly in children. The large historical cohorts that made it possible to study late effects mainly concern patients treated between 1950 and 2000, and thus little or not at all exposed to these new modalities. There is therefore a crucial need to build and exploit contemporary cohorts to assess the iatrogenic risks associated with these innovations.
In this perspective, the EpiRad team initiated the Horus cohort, based on data from four Comprehensive Cancer Centres (Gustave Roussy (GR), Institut Curie (IC), Centre Léon Bérard (CLB), Centre Oscar Lambret (COL)), equipped with the ConSoRe system. This system allows semi-automated, structured collection of clinical data from multiple sources, facilitating reconstruction of patients' oncological and medical history. These data will be enriched by linkage with the SNDS.
Exploiting the SNDS medico-administrative databases offers major prospects for the long-term follow-up of childhood cancer survivors and the identification of late sequelae in the general population. However, these data are subject to misclassification of health events, which may introduce major biases into epidemiological analyses [7]. In this context, developing algorithms to identify sequelae from the SNDS is a central methodological challenge, within the broader field of real-world data exploitation and health data phenotyping [8-10]. These algorithms must rely on complex care trajectories reconstructed from sequences of events (hospitalisations, prescriptions, procedures), and be validated against sources with a clinical gold standard (medical records, self-reported then validated questionnaires, and long-term follow-up consultations), such as the FCCSS cohort.
The main objective of this thesis is to develop and validate algorithms to identify late sequelae in survivors of childhood and adolescent and young adult (AYA) cancers.
Secondary objectives
1) Develop and evaluate the performance of SNDS algorithms for various sequelae (second cancers, cardiovascular, neurological, endocrine), using care-trajectory modelling and sequence analysis approaches.
2) Study the iatrogenic effects of innovative drug therapies (compared with conventional treatments) in the Horus cohort.
3) Generalise the results to all childhood and AYA cancer survivors in France using the exhaustive SNDS database. PLANNED WORK - METHODS & RESOURCES
Axis 1: Construction and validation of algorithms to identify long-term sequelae in the SNDS
Materials and methods
Data from the FCCSS cohort, linked with SNDS data, will be used. The FCCSS is a retrospective cohort including 7,670 patients diagnosed with a solid tumour or lymphoma before the age of 21 and before the year 2000 in five French cancer centres. For each individual, detailed data were collected from medical records (clinical characteristics, treatments, demographic data). Long-term follow-up relies on several sources: medical records, self-reported questionnaires, long-term clinical follow-up consultations and SNDS data. With a median follow-up of about 32 years, the FCCSS cohort has made it possible to identify a large number of clinically validated health events, including second cancers, cardiovascular diseases, hearing disorders, diabetes, cerebrovascular diseases and renal failure.
A subsample of 5,589 participants with validated clinical data will be used as the reference (gold standard) for algorithm development. The aim will be to estimate, from SNDS data (ICD-10 codes, CCAM procedures, drug dispensing, hospitalisations), the probability that an observed event corresponds to a true clinical event.
Statistical analyses
Individual care trajectories will be reconstructed and analysed using sequence analysis, trajectory clustering, Markov models, and machine learning approaches [11,12]. Algorithm performance will be evaluated using the area under the ROC curve (AUC), as well as standard classification indicators (sensitivity, specificity, predictive values).
Axis 2: Study of iatrogenic effects of innovative drug therapies (compared with conventional treatments) in the Horus cohort
Materials and methods
In each centre, the ConSoRe system enabled the collection of demographic, clinical and therapeutic data. The HORUS-ConSoRe cohort currently comprises nearly 20,000 children, adolescents and young adults: Gustave Roussy (n = 7,941), Institut Curie (n = 6,813), Centre Léon Bérard (n ~ 4,000) and Centre Oscar Lambret (n ~ 2,000). After pseudonymisation, the data are centralised by the EpiRad team. Linkage with the SNDS will be performed using the national registry number (NIR), according to a schedule with a first linkage in 2026 and triennial updates until 2034 (CNIL authorisation no. 925155). This linkage will allow longitudinal follow-up of patients, including the occurrence of deaths, secondary events and late complications.
Exposure to innovative drug therapies (targeted therapies, immunotherapies, cell therapies) will be identified from hospital data (ConSoRe) and community care data (SNDS). A nomenclature of innovative treatments will be built by combining information from clinical and medico-administrative databases (ATC, UCD, CIP, GHS, CCAM codes), allowing fine characterisation of exposures. The algorithms developed in Axis 1 will be applied to this cohort to identify late sequelae (second cancers, cardiovascular, neurological, endocrine diseases).
Data analysis
The analysis will rely on descriptive and analytical approaches aimed at estimating the association between exposure to innovative therapies and the occurrence of long-term iatrogenic events.
Survival models will be used, including Cox proportional hazards models to estimate adjusted hazard ratios, as well as Fine-Gray models to account for competing risks (notably death) [13]. In addition to propensity-score-based methods (IPTW, matching), a target trial emulation approach will be considered, explicitly defining the hypothetical protocol (eligibility criteria, treatment strategies, assignment, time zero, follow-up and outcomes), in order to limit confounding by indication, immortal time bias and selection bias, and to strengthen the causal interpretation of estimates [14]. Interaction analyses will be conducted to explore combined effects between innovative therapies and conventional treatments (chemotherapy, radiotherapy, surgery).
Finally, sensitivity analyses will be carried out to assess the robustness of results to modelling assumptions and residual biases (notably misclassification of exposures and events).
Axis 3: Generalisation of results at the national level to childhood and AYA cancer survivors in France using the exhaustive SNDS database
Rationale and issues
Although the HORUS cohort offers a high level of clinical detail, it remains subject to selection bias, related to the recruitment of patients treated in expert centres, often with more complex forms of disease or specific biological characteristics. This bias is likely to affect the estimation of exposure prevalence and of the risks associated with innovative therapies. To improve the external validity of the results, a generalisation at the national level will be carried out using exhaustive SNDS data.
Materials and methods
The HORUS-SNDS cohort will include all children, adolescents and young adults diagnosed with cancer before the age of 26 between 2006 and 2024 in France (mainland France and Corsica), identified in the SNDS (protocol currently being validated by Inserm's IT department (DSI) for access to the exhaustive SNDS database).
Inclusion criteria will be: a cancer diagnosis identified in the SNDS, age at diagnosis strictly below 26, affiliation to the French social security system, and residence in mainland France or Corsica. Excluded will be individuals whose age at diagnosis is 26 or above, those not affiliated to the French healthcare system, and patients who died before the start of the observation period, set at 2006.
Exposures to innovative drug therapies will be identified from the various sources available in the SNDS, using codes for drugs (ATC, CIP, UCD), medical procedures (CCAM), hospital stays (GHS) and medical devices (LPP), according to a nomenclature harmonised with that developed in Axis 2. Late sequelae will then be identified using the algorithms previously developed and validated in Axis 1.
Data analysis
The analyses will follow an approach similar to that developed in Axis 2, with adaptations related to the nature of medico-administrative data. The main objectives will be:
- to estimate the prevalence and incidence of exposure to innovative therapies at the national level,
- to assess the risks of late sequelae associated with these exposures,
- and to compare these estimates with those from the HORUS cohort in order to identify possible discrepancies due to selection bias.
Survival models (Cox, Fine-Gray) will be used, with adjustment for variables available in the SNDS (age, sex, comorbidities). Finally, transportability and generalisation methods may be used to combine the respective strengths of the two data sources (clinical richness vs. population exhaustiveness).