Lymphoma is one of the top ten most frequently diagnosed cancers worldwide, with an estimated 89,380 new cases projected in the United States in 2023. The five-year mortality rate exceeds 25%, and the disease was expected to claim more than 21,080 lives that same year. Phase 1 clinical trials occupy a critical role in the drug development pipeline, representing the first stage of human testing for investigational agents and translating years of preclinical laboratory research into actionable clinical solutions. Without accurate duration forecasting, planning these trials is largely guesswork that consumes enormous institutional resources.
The forecasting gap: Clinical trial planning tools have historically focused almost entirely on calculating sample size, the number of participants needed for adequate statistical power. Widely used calculators such as those from the NIH (GRT Calculator), OpenEpi, and ClinCalc address sample size but provide no estimate of how long a trial will actually take to complete. This gap has real financial and logistical consequences: nearly 85% of clinical trials experience setbacks, driven by strategic challenges, commercial barriers, operational problems, and unmanageable toxicity. Delays in clinical drug development cost sponsors an estimated USD 600,000 to USD 8 million per day.
Why machine learning: Prior statistical approaches to trial duration estimation, such as the parametric model by Rubinstein et al. based on exponential death time assumptions, depend heavily on restrictive distributional assumptions and focus narrowly on adverse event endpoints. They do not account for variables such as enrollment numbers, trial locations, intervention types, sponsor characteristics, and study design patterns, all of which collectively influence how long a trial runs. Machine learning offers a fundamentally different approach: it can ingest large volumes of actual historical trial records alongside dozens of trial-specific features and learn non-linear predictive relationships that standard regression models cannot capture.
This paper introduces a binary machine learning model trained on 1,089 completed Phase 1 lymphoma trials sourced from clinicaltrials.gov, designed to predict whether a given trial will complete within or beyond the average duration of approximately five years (1,788 days). Rather than attempting to predict exact durations, the binary framing provides actionable, interpretable outputs that are less susceptible to distortion from outliers while still offering planners a clear basis for resource allocation decisions.
The dataset was sourced exclusively from clinicaltrials.gov, the National Library of Medicine's global registry for clinical research studies. The authors restricted their query to Phase 1 trials tagged with "Lymphoma" that had started before 2023 and were marked as "Completed," yielding an initial collection of 1,231 studies. Trials initiated in 2023 were deliberately excluded to avoid seasonality artifacts, since only partial-year data were available at the time of data collection (July 2023). The duration variable was computed as the interval between the trial's recorded Start Date and its Completion Date.
Preprocessing steps: Trials with missing start or completion dates were removed from the lymphoma dataset, reducing it from 1,231 to 1,089 usable studies. The data were split 80/20 for training and testing (871 training, 218 testing), with 5-fold cross-validation applied to the training set for hyperparameter tuning. Missing values in numerical variables such as enrollment were imputed using the column mean. Categorical variables with missing values were treated as a distinct separate category rather than imputed, preserving the information that the field was absent.
Target variable construction: The average duration of the 1,089 Phase 1 lymphoma trials was 1,788 days, approximately five years. A binary target variable was derived from this benchmark: trials exceeding 1,788 days were labeled class 1 (longer than average), while those at or below were labeled class 0 (within average). Approximately 40% of trials in both the training and testing sets fell into the longer-than-average class, producing a moderately imbalanced dataset that nonetheless allowed standard evaluation metrics to remain interpretable without resampling.
External validation dataset: For out-of-sample generalization testing, the authors assembled a parallel dataset of 907 completed Phase 1 lung cancer trials from clinicaltrials.gov, preprocessed in the same manner and reduced to 840 usable records after date exclusions. Lung cancer was selected because it shares clinical trial design conventions, regulatory requirements, and common endpoint structures with lymphoma, making it a meaningful test of cross-cancer generalizability rather than a trivial sanity check.
Through exploratory data analysis, the authors identified a set of trial characteristics that significantly predict completion time. Enrollment size showed one of the strongest associations: trials with 0-20 participants averaged 1,417 days to complete, while those enrolling 61 or more participants averaged 2,218 days, a difference of roughly 1.6 times that was statistically significant at the 95% confidence level. Larger trials require more time to recruit, treat, and follow up patients, making enrollment a naturally strong predictor of logistical burden.
Sponsor type and trial complexity: Industry-led trials (n = 510) completed significantly faster than non-industry-led trials (n = 579), with mean durations of 1,414 days versus 2,118 days respectively (p less than 0.05). This difference likely reflects the operational infrastructure and project management resources that pharmaceutical companies bring to trial execution. Trial complexity also matters: those examining more than three distinct medical conditions lasted an average of 215 days longer than those with fewer conditions (1,929 days versus 1,714 days), and trials involving more than one intervention took an average of 248 days longer than single-intervention trials (1,909 days versus 1,661 days).
NLP-assisted feature creation: For columns containing substantial free text, particularly "Outcome Measures" and "Sponsor/Collaborators," the authors used the spaCy natural language processing library to compute semantic similarity between terms. Words with similarity scores above 0.8 were grouped using a Disjoint Set Union (DSU) algorithm, enabling normalization of synonymous terms such as "adverse events," "adverse reactions," and "AEs" into a single outcome measure category. This approach avoids reducing complex text fields to simple string matching and allows the model to recognize that semantically equivalent but textually different entries represent the same clinical concept.
Counterintuitive findings: Trials that included adverse event outcome measures actually completed faster on average (1,537 days with AE measurement versus 1,919 days without, p less than 0.05). This may reflect that safety-focused Phase 1 trials are often more tightly scoped. Conversely, NCI-sponsored trials had a mean duration of 2,246 days compared to 1,648 days for non-NCI trials, and trials involving biological interventions ran longer than those without. Following iterative feature selection, 30 features were incorporated into the final models, ranked by Gini Gain importance. The top features by importance were enrollment count, industry leadership status, location count, outcome measures count, condition count, and intervention count.
Using Python 3.9.7, the authors trained and evaluated eight distinct machine learning classifiers on the lymphoma dataset. The selection was guided by prior work on oncology clinical trial prediction and chosen to span the major families of supervised learning approaches. The eight models were: Logistic Regression (LR), K-Nearest Neighbors (KNN), Decision Tree (DT), Random Forest (RF), XGBoost (XGB), Linear Discriminant Analysis (LDA), Gaussian Naive Bayes (Gaussian NB), and Multi-Layer Perceptron (MLP). This breadth allowed systematic comparison across linear, instance-based, tree-based, probabilistic, and neural network paradigms.
Hyperparameter tuning: Each model was refined using GridSearchCV (GSCV) from the Scikit-Learn library, which exhaustively searches a defined hyperparameter grid using 5-fold cross-validation to identify the best-performing configuration. This process prevents the naive mistake of reporting performance on the same data used to tune the model, providing more reliable estimates of generalization. For Random Forest, the grid spanned tree counts of 50 to 500, maximum depths of none, 10, 20, and 30, minimum sample splits of 2, 5, and 10, and both bootstrap options. The optimal configuration used a max depth of 20, min samples split of 10, 100 trees, and no bootstrap (random_state = 42).
Evaluation metrics: Performance was assessed using accuracy, ROC-AUC, precision, recall, and F1-score, all computed on the held-out 20% test set as well as during 5-fold cross-validation. Accuracy measures the fraction of correct predictions across both classes. The AUC of the ROC curve integrates classifier performance across all decision thresholds, with values closer to 1.0 indicating better discrimination between above-average and below-average duration trials. Precision captures the fraction of trials predicted to exceed five years that actually did so. Recall captures the fraction of trials that actually exceeded five years and were correctly identified. The F1-score harmonically averages precision and recall into a single metric that balances the trade-off between the two.
Model families compared: Logistic Regression and LDA assume linear decision boundaries, with LDA additionally assuming class-conditional Gaussian distributions with equal covariances. KNN classifies based on the majority label among the k nearest training examples, making it highly sensitive to feature scaling and computationally expensive at inference. Decision Trees are fully expressive but prone to overfitting the training data. Random Forest and XGBoost both use ensembles of decision trees: Random Forest aggregates trees trained on bootstrap samples using averaging (bagging), while XGBoost builds trees sequentially, each one correcting the residuals of its predecessors (boosting). Gaussian NB applies Bayes' theorem assuming feature independence, which produces high recall but typically compromises precision. MLP uses fully connected neural network layers, offering theoretical flexibility to model complex non-linearities but lacking transparency.
Across 5-fold cross-validation on the training set, XGBoost and Random Forest led all eight classifiers. XGBoost achieved an accuracy of 74.42% (plus or minus 3.84%) and an ROC-AUC of 78.54% (plus or minus 3.89%), while Random Forest followed closely with an accuracy of 73.71% (plus or minus 3.89%) and ROC-AUC of 77.55% (plus or minus 4.18%). Logistic Regression and LDA produced comparable accuracy levels of 71.18% and 70.72% with ROC-AUC values of 77.60% and 75.67%, respectively. The Multi-Layer Perceptron performed more poorly, achieving 67.17% accuracy and 70.71% ROC-AUC, with a notably high standard deviation in recall (49.14% plus or minus 9.84%), suggesting instability across cross-validation folds. Gaussian Naive Bayes achieved a striking recall of 90.86% but at the cost of only 52.93% overall accuracy, reflecting its tendency to predict the positive class too liberally when feature independence assumptions are violated.
Test set results for top four models: On the held-out 218-trial test set, Random Forest was the top performer with accuracy of 0.7248, ROC-AUC of 0.7677, precision of 0.675, recall of 0.6136, and F1-score of 0.6429. Notably, XGBoost, which ranked first in cross-validation, dropped to third on the test set with accuracy of 0.6881 and ROC-AUC of 0.7574, a pattern consistent with overfitting to noise in the training data. LDA moved into second position with test accuracy of 0.6927 and ROC-AUC of 0.7319, likely because its distributional assumptions provided a useful inductive bias that promoted generalization. Logistic Regression maintained fourth position with accuracy of 0.6422 and ROC-AUC of 0.7281.
Statistical significance testing: To formally verify that Random Forest's accuracy advantage over the other three top models was not due to chance, the authors conducted chi-square tests comparing the number of correctly predicted trials between Random Forest (158 of 218) and each competitor on the test set. LDA correctly predicted 151 of 218 trials (p = 1.45 x 10^-17 versus RF), XGBoost correctly predicted 150 of 218 trials (p = 4.35 x 10^-30 versus RF), and Logistic Regression correctly predicted 140 of 218 trials (p = 2.77 x 10^-21 versus RF). All three p-values were far below the 0.05 threshold, confirming that Random Forest's superiority was statistically significant at the 95% confidence interval.
Calibration curve analysis: The authors further assessed model reliability using calibration curves, which plot the observed fraction of trials exceeding five years against the model's predicted probability for each probability bin. Within the 0 to 0.6 predicted probability range, all four top models closely aligned with the ideal diagonal calibration line, with XGBoost showing the most deviation. Beyond 0.6 predicted probability, all models showed more deviation, but Random Forest remained the closest to perfect calibration and uniquely returned toward the calibration line in the high-probability region (above 0.8), while the other models continued to diverge.
A central contribution of this paper is the demonstration that the Random Forest model's predicted probability output carries information beyond a simple binary yes/no classification. After training, the model assigns each trial a probability score between 0 and 1 reflecting the estimated likelihood that the trial will exceed five years. The authors divided the 218 test-set trials into five quantile groups based on this predicted probability and then measured the actual average duration of trials within each group, along with 95% confidence intervals.
Quantile group results: The five quantile groups showed a clear monotonic increasing relationship between predicted probability and actual average duration. Trials in the first quantile (predicted probability 0 to 0.1624) had an average actual duration of 1,140 days (about 3.12 years), with a 95% CI of 935 to 1,345 days. Trials in the second quantile (0.1624 to 0.3039) averaged 1,541 days, with a CI of 1,235 to 1,847 days. The third quantile (0.3039 to 0.4697) averaged 1,799 days (CI: 1,557 to 2,041 days). The fourth quantile (0.4697 to 0.6291) averaged 2,150 days (CI: 1,730 to 2,569 days). Trials in the highest quintile (0.6291 to 1.0) averaged 2,352 days, or approximately 6.44 years, with a CI of 2,005 to 2,699 days.
Practical significance: This graduated probability-to-duration mapping transforms the model from a blunt binary classifier into a more nuanced planning instrument. A trial assigned a predicted probability of 0.05 can be told with reasonable confidence it will likely complete in the range of 935 to 1,345 days, whereas a trial assigned a predicted probability of 0.75 should be planned for a range closer to 2,005 to 2,699 days. This precision goes beyond the binary output and provides stakeholders with an expected duration range, a 95% confidence interval for that range, and a probability score that itself conveys relative timing risk, enabling tiered resource planning decisions.
Confidence interval behavior: Consistent with the uncertainty inherent in clinical trial prediction, the upper bounds of the 95% CIs tended to narrow with increasing predicted probability, while lower bounds followed an inverse pattern. The widest relative intervals appeared at intermediate probability ranges (quantiles 3 and 4), reflecting the greatest uncertainty about whether these borderline trials will fall on the shorter or longer side of the five-year threshold. This heteroskedastic pattern is epidemiologically plausible and provides honest communication of prediction precision to planners.
To assess whether the 871-trial training set was appropriately sized for the task, the authors performed a learning curve analysis, training the Random Forest model on progressively increasing fractions of the training data (from 20% to 100%) and measuring test set accuracy and ROC-AUC at each increment. The analysis revealed a positive but diminishing returns relationship: accuracy gains were most pronounced between 20% and 60% training data usage, but the rate of improvement tapered substantially beyond the 60% mark. At the full 871-trial training set, the model reached its peak accuracy of 0.7248. The highest ROC-AUC was recorded at 60% data usage, but the difference compared to the full dataset was slim, confirming that the current dataset size is close to optimal for this task and that collecting substantially more data from clinicaltrials.gov is unlikely to produce a meaningful accuracy boost.
External validation on lung cancer: The fully trained Random Forest model was applied without retraining to the 840-trial lung cancer external validation set, treating it as a completely held-out population the model had never encountered. The Random Forest classifier demonstrated the highest performance among all eight models on this external set as well, achieving an accuracy of 0.7405 and an ROC-AUC of 0.7701, both marginally higher than on the lymphoma test set. Logistic Regression and LDA followed with accuracies of 0.7321 and 0.7310 and ROC-AUC values of 0.7671 and 0.7647 respectively. XGBoost ranked fourth with accuracy 0.725 and ROC-AUC 0.7632.
Interpretation of cross-cancer generalizability: The fact that the lymphoma-trained model performed at least as well on lung cancer trials as it did on lymphoma trials suggests that the features driving Phase 1 trial duration are largely shared across cancer types. Enrollment size, sponsor type, number of conditions, intervention complexity, and outcome measure structure appear to be generalizable predictors of Phase 1 trial logistics regardless of the specific oncological indication. The consistent ranking of Random Forest as the top model on both datasets further supports this conclusion.
The lung cancer validation also demonstrated the same probability-to-duration monotonic increasing trend observed in the lymphoma analysis, with trials assigned higher predicted probabilities consistently averaging longer actual durations. This cross-cancer replication of the quantile finding suggests the probability output is a reliable continuous risk score, not an artifact of the lymphoma-specific training data.
Dataset scope and selection bias: The dataset was drawn exclusively from clinicaltrials.gov, which, while the most comprehensive publicly available global registry, does not capture all Phase 1 lymphoma trials conducted worldwide, particularly those registered in regional databases such as the EU Clinical Trials Register or the Chinese Clinical Trials Registry. This may introduce geographic and institutional biases. The decision to exclude trials with missing start or completion dates, while methodologically necessary, could inadvertently remove trials with distinctive patterns, such as unusually long trials that were administratively delayed in date reporting.
Imputation limitations: Mean imputation was used for the numerical enrollment variable where values were missing. This common approach reduces the variance of the imputed variable and may attenuate the predictive strength of enrollment in the model. More sophisticated imputation strategies, such as multiple imputation by chained equations (MICE) or k-nearest neighbor imputation, could potentially preserve more of the underlying distribution and improve model performance, particularly for trials where enrollment was systematically missing for structural rather than random reasons.
Model interpretability: Random Forest, while more interpretable than deep neural networks, is still an ensemble model whose predictions arise from hundreds of decision trees. The Gini Gain importance scores provide a useful ranking of feature contributions but do not directly explain how each feature combination interacts to produce a prediction for a specific trial. Future work incorporating SHAP (SHapley Additive exPlanations) values would allow individual trial-level explanation of predictions, providing trial planners with a clear statement of which specific design choices are most strongly driving the model's output for their particular protocol.
Practical impact and future extensions: The authors note that clinical trial delays cost USD 600,000 to USD 8 million daily, and that approximately 10% of clinical drug development failures are attributable to poor strategic planning. By providing duration estimates with confidence intervals, this model has the potential to directly inform trial prioritization, staggered start dates, and personnel allocation decisions at research organizations. Future extensions could incorporate trial-specific molecular data on the investigational agents, expand the model to Phase 2 and Phase 3 trials, and validate predictions against trials in additional cancer types such as breast cancer, colorectal cancer, and myeloma. Integration with electronic research management systems could allow real-time updating of duration predictions as trials progress and new data accumulate.