Many cancer patients undergoing immune-checkpoint inhibition (ICI) experience early progression (EP), defined as progression at the first imaging evaluation. Circulating cell-free DNA (cfDNA) fragmentation profiles, especially fragment size distribution, offer a promising non-invasive method for assessing treatment response independently of a specific molecular target, cancer type, and treatment. However, only few computational modeling studies investigated the underlying biological mechanisms of cfDNA dynamics during treatment monitoring. The SChISM (Size CfDNA Immunotherapies Signature Monitoring) clinical study focuses on monitoring the plasmatic cfDNA size profile of the fragmentome in 139 advanced pan-cancer patients undergoing ICI treatment. The aim is to early manage therapeutic strategy before ICI-related progression or toxicity occurs. We performed statistical and survival biomarker analysis aimed to predict EP and progression-free survival (PFS) using pre-treatment data. Plasmatic fragmentome-derived metrics, including concentration, size distribution of first and second peaks, and specific size ranges (p = 11 variables), were evaluated alongside standard clinical markers such as ECOG performance status, age, and pathology. The relative quantity of fragments over 1650 bp exhibited the highest discriminatory power for EP (median area under ROC curve on the test set = 0.7, 95% CI: 0.59-0.80). A lower relative quantity of these long fragments correlated with EP and shorter PFS in both the univariate and multivariate settings. Following the development of empirical models capturing inter- and intra-patient variability in cfDNA metrics, a system of differential equations was constructed to individually describe cfDNA kinetics. Subsequently, a mechanistic model integrating cfDNA kinetics with longitudinal biological markers and tumor size imaging was developed to elucidate and comprehend the temporal dynamics of cfDNA quantitative profiles over time. Using population data, the model was calibrated within the nonlinear mixed effects statistical framework. Future perspective aim to employ machine learning models to predict EP, PFS or overall survival by integrating dynamic parameters with baseline variables.