In a retrospective cohort study of ninety patients with advanced soft tissue sarcoma treated at LMU Hospital Munich between January 2021 and December 2025, researchers applied machine learning and multivariable survival modeling to uncover key determinants of overall survival beyond second-line therapy. The investigators utilized extreme gradient boosting to identify clinical features predicting survival of at least twelve months following third-line treatment, subsequently testing top variables in multivariable Cox models. In the eighty-eight-patient third-line cohort, failure to achieve at least stable disease on second-line therapy independently predicted worse overall survival, generating a hazard ratio of 2.31 (p = 0.005). Conversely, liposarcoma or leiomyosarcoma histology was associated with significantly improved survival (HR 0.37, p = 0.002), as was a time to progression of twelve months or longer during first-line therapy (HR 0.43, p = 0.007). Exploratory sequencing analyses further suggested distinct timing effects among standard systemic therapies, showing differing survival associations across successive treatment lines.
The clinical rationale for this investigation stems from the substantial uncertainty characterizing later-line systemic therapy in advanced soft tissue sarcoma. While anthracycline combinations such as doxorubicin-based regimens represent the established standard of care in first-line therapy, disease progression occurs in most patients, necessitating subsequent lines of treatment. Beyond the initial setting, clinical decisions are largely guided by histologic subtype, prior therapy exposure, and the pragmatic objective of achieving disease stabilization rather than curative outcomes. However, prospective randomized trials predominantly focus on single lines of therapy within strictly defined populations, offering minimal real-world insight into patient trajectories across successive third- and fourth-line interventions. Comparative data regarding whether the specific order and timing of active agents influence long-term overall survival remain scarce due to small subgroup sizes in clinical trials. Consequently, oncologists face persistent challenges when selecting and ordering available therapies, making data-driven modeling of retrospective cohorts an important approach for clarifying late-line prognostic determinants.
Prior clinical studies have established the efficacy boundaries of several systemic agents commonly administered in advanced pretreated sarcoma, though reported survival benefits remain generally modest. In non-adipocytic sarcoma following anthracycline therapy, the oral multi-kinase inhibitor pazopanib prolonged median progression-free survival to 4.6 months compared with 1.6 months for placebo, without improving overall survival (median 12.5 versus 10.7 months), demonstrating notable activity in leiomyosarcoma and synovial sarcoma. Gemcitabine-based regimens, particularly gemcitabine combined with docetaxel, have documented median progression-free survival of approximately 3 to 6 months and median overall survival of 10 to 18 months, achieving higher response rates in leiomyosarcoma. For liposarcoma and leiomyosarcoma, trabectedin demonstrated superior progression-free survival over dacarbazine (median 3.3 versus 2.3 months) without overall survival extension (median 13.9 versus 11.8 months), while retaining disease control across later lines. Other regimens, including eribulin, dacarbazine, and temozolomide, exhibit variable activity, yet median survival frequently drops to approximately 6.4 months once standard therapies are exhausted.
The study analyzed ninety consecutive adult patients with histologically confirmed soft tissue sarcoma who received at least three lines of systemic therapy at LMU Hospital Munich between January 2021 and December 2025. The research protocol received institutional approval under protocol number 26-0142, with all histopathological diagnoses independently confirmed by a dedicated sarcoma pathologist. Within the cohort, median patient age at initial diagnosis was 50 years, with an interquartile range of 39 to 62 years, and 54 percent were female. High-grade histology predominated, with 52 percent classified as grade 3 and 43 percent as grade 2 according to FNCLCC criteria. Primary tumors were visceral in 48 percent and extremity-based in 33 percent of patients. Leiomyosarcoma represented the most frequent subtype at 41 percent, followed by synovial sarcoma at 16 percent, liposarcoma at 12 percent, undifferentiated pleomorphic sarcoma at 10 percent, and rare histologies at 21 percent. While 30.6 percent presented with synchronous metastases, 69.4 percent developed metastases metachronously, with 41 percent exhibiting oligometastatic disease of five or fewer lesions.
Across successive treatment lines, clinical regimens diversified considerably from standardized front-line regimens to heterogeneous salvage selections. First-line systemic treatment was dominated by anthracycline-based protocols in 92 percent of patients, primarily doxorubicin plus ifosfamide (62.2 percent), doxorubicin plus dacarbazine (20.0 percent), or doxorubicin monotherapy (8.9 percent). In second-line therapy, trabectedin was the most frequent choice (25.6 percent), where radiological evaluation demonstrated progressive disease in 48 percent, stable disease in 39 percent, partial response in 12 percent, and no complete responses. Third-line therapy most frequently involved pazopanib (28.9 percent) and trabectedin (26.7 percent). By the fourth line, trabectedin (27.6 percent) and pazopanib (24.1 percent) remained common, but choices broadened widely. In the fifth line, therapeutic diversification reached its peak, with unlisted therapies accounting for 42.3 percent of regimens, including dacarbazine monotherapy, gemcitabine combinations, high-dose ifosfamide, and experimental protocols in clinical trials. This transition illustrates the progressive fragmentation of clinical strategies once patients exhaust initial lines of treatment.
To identify clinical predictors of overall survival of at least twelve months following third-line therapy initiation, the researchers implemented an extreme gradient boosting model using stratified five-fold cross-validation. Extreme gradient boosting constructs an ensemble of decision trees sequentially, with each tree trained to correct the predictive errors of preceding trees, enabling the model to capture non-linear relationships and clinical interactions. Candidate variables included patient age at metastatic diagnosis dichotomized at 55 years, sex, histological subtype (liposarcoma or leiomyosarcoma versus other), metastatic burden (five or fewer versus more than five lesions), primary tumor site, time to first metastasis diagnosis, time to progression after first-line therapy (twelve months or longer versus less), RECIST version 1.1 radiological response to second-line therapy, third-line trabectedin use, third-line local interventions, and prior maintenance therapy. Near-zero-variance predictors were filtered out, and hyperparameters including tree depth, boosting rounds, and learning rate were systematically tuned to maximize cross-validated area under the curve.
The finalized extreme gradient boosting model achieved a cross-validated area under the receiver operating characteristic curve of 0.69, indicating reasonable discriminative ability for predicting one-year survival from third-line therapy start. Gain-based importance metrics identified best radiological response to second-line therapy—specifically distinguishing disease progression from disease control comprising partial response or stable disease—as the single most influential predictor in the model. Histological subtype, contrasting liposarcoma and leiomyosarcoma against other histological variants, emerged as the second strongest predictor, followed by time to progression on first-line therapy of twelve months or longer. Additional influential predictors included baseline metastatic lesion burden, patient age, and sex. Conversely, treatment-specific factors, including third-line trabectedin use, time to initial metastatic recurrence, prior maintenance therapy, and metastasis-directed local interventions, demonstrated comparatively modest gains, indicating that underlying biological responsiveness and disease tempo were far more decisive than specific local procedural interventions. These algorithmic findings provided the empirical foundation for subsequent multivariable survival regressions.
To establish clinically interpretable hazard ratios and account for event rates, top-ranking features from the machine learning model were evaluated in multivariable Cox regressions. From third-line initiation, median progression-free survival was 3.1 months (95% CI 2.8 to 5.0) and median overall survival was 20.9 months (95% CI 16.4 to 25.0). In eighty-eight third-line patients, second-line progression independently doubled mortality hazard (HR 2.31, 95% CI 1.28 to 4.18, p = 0.005). Favorable survival was associated with liposarcoma or leiomyosarcoma histology (HR 0.37, 95% CI 0.20 to 0.70, p = 0.002) and first-line progression-free interval of twelve months or longer (HR 0.43, 95% CI 0.23 to 0.79, p = 0.007). In fifty-seven fourth-line patients, median progression-free survival was 4.0 months and median overall survival dropped to 10.8 months. Second-line progression remained detrimental (HR 2.06, 95% CI 1.03 to 4.09, p = 0.040) and liposarcoma or leiomyosarcoma maintained significant benefit (HR 0.28, 95% CI 0.13 to 0.60, p = 0.001), while first-line disease control trended protectively (HR 0.52, 95% CI 0.25 to 1.06, p = 0.074).
The researchers also conducted exploratory sequencing analyses to determine whether the timing and order of standard late-line regimens affected overall survival. To minimize survivor and immortal-time bias, this evaluation was restricted to patients who received at least four lines of systemic therapy, tracking survival from second-line therapy start. Distinct drug-specific survival patterns emerged across trabectedin, pazopanib, and gemcitabine plus docetaxel. Trabectedin administration demonstrated its longest overall survival when utilized in the fourth line rather than earlier lines (p = 0.023), suggesting sustained therapeutic activity in heavily pretreated patients. In contrast, pazopanib showed more favorable survival when administered earlier in the sequence, specifically during second-line therapy, with diminishing benefit when introduced in subsequent lines (p = 0.022). Gemcitabine combined with docetaxel produced overlapping survival curves across second-, third-, and fourth-line administration with no statistically significant differences (p = 0.12), indicating line-independent cytotoxic efficacy across sequential therapeutic stages. These findings offer hypothesis-generating insights into agent-specific timing.
The authors suggested that trabectedin’s sustained late-line activity may stem from its distinct mechanism of action, absence of complete cross-resistance, and efficacy in responsive histologic subtypes. In contrast, pazopanib’s earlier benefit aligns with lower pretreatment burden and declining antiangiogenic efficacy in heavily pretreated, resistant disease. Several important methodological limitations qualify these retrospective observations. Because treatment allocation was non-randomized, selection bias and patient performance status inevitably influenced regimen choice across lines. The limited sample size of ninety patients constrained statistical power for detecting modest differences, and the predominance of leiomyosarcomas along with visceral and retroperitoneal tumors at LMU Hospital Munich limits generalizability to extremity-predominant sarcoma populations. Furthermore, data-driven machine learning in moderately sized cohorts carries inherent risks of overfitting and winner’s curse. The investigators therefore concluded that their findings provide a foundational framework for understanding later-line sarcoma trajectories, emphasizing that identified prognostic factors and sequencing patterns require confirmation in prospective, independent cohorts.
Subject of Research: Machine learning survival modeling and systemic therapy sequencing in advanced soft tissue sarcoma beyond second-line treatment
Article Title: Machine Learning-Guided Survival Prediction and Treatment Sequencing in Advanced Soft Tissue Sarcoma Beyond Second-Line Therapy: A Retrospective Cohort Study
Article References: Hoberger, M., Di Gioia, D., Zuber, R. L., Völkl, M., Güler, S. E., Jurinovic, V., Albertsmeier, M., Klein, A., Dürr, H. R., Schmidt–Hegemann, N.-S., Knösel, T., Kunz, W. G., von Bergwelt–Baildon, M., Lindner, L. H., Burkhard–Meier, A., & Berclaz, L. M. (2026). Machine Learning-Guided Survival Prediction and Treatment Sequencing in Advanced Soft Tissue Sarcoma Beyond Second-Line Therapy: A Retrospective Cohort Study. Oncology and Therapy, 14(3), 967-981. https://doi.org/10.1007/s40487-026-00446-7
Image Credits: AI Generated
DOI: 10.1007/s40487-026-00446-7
Keywords: Soft Tissue Sarcoma, Machine Learning, XGBoost, Overall Survival, Progression-Free Survival, Second-Line Therapy, Third-Line Therapy, Trabectedin, Pazopanib, Gemcitabine, Docetaxel, Leiomyosarcoma, Liposarcoma, Cox Proportional Hazards Model, Therapy Sequencing, Retrospective Cohort Study, Oncology and Therapy, LMU Hospital Munich
News Source: Teresa Odom. (November 19, 2026). Machine Learning Unveils Key Survival Factors in Advanced Soft Tissue Sarcoma. Scienmag.



