Technical Summary: Explainable Machine Learning for Clinical Phenotyping and Mortality Prediction in Critically Ill Patients with Cancer

1. Problem Statement

Critically ill patients with cancer present a unique clinical challenge due to their high heterogeneity and complex pathophysiology, which often involves the interplay of tumor burden, treatment-related toxicities, and physiological reserve depletion. Traditional severity scoring systems (e.g., SOFA, APACHE) rely on linear assumptions and have demonstrated limited prognostic value in this specific population, where short-term mortality remains high (30%–50%). Furthermore, the "black box" nature of advanced machine learning models has hindered their clinical translation, as clinicians require interpretable decision-making logic to trust predictive tools. There is a critical need for data-driven approaches that can identify latent clinical subgroups and provide explainable, early mortality predictions to support risk stratification and avoid futile treatments.

2. Methodology

Data Sources and Cohort Selection

This multicenter, retrospective observational study utilized two large, de-identified public databases:

• Derivation Cohort: MIMIC-IV (Beth Israel Deaconess Medical Center, 2008–2019), containing 14,573 adult patients with solid tumors or hematological malignancies.
• External Validation Cohort: eICU Collaborative Research Database (over 200 US hospitals), containing 10,301 similar patients.
• Inclusion/Exclusion: Patients were included if they were ≥18 years old with a confirmed cancer diagnosis. Exclusions included non-cancer ICU admissions, repeat admissions during the same hospitalization, and cases with >30% missing core data.

Data Preprocessing

• Feature Selection: 28 routinely available variables from the first 24 hours of ICU admission were used, including demographics, cancer categories, vital signs, laboratory results, and therapeutic interventions.
• Exclusion of Composite Scores: Traditional severity scores (SOFA, APACHE IV) were extracted only for baseline comparison and were strictly excluded from the machine learning feature space to ensure the model relied on raw physiological data.
• Handling Missingness: Missing values for continuous variables were imputed using the median from the derivation cohort (MIMIC-IV) to prevent data leakage. The XGBoost algorithm's built-in handling of missing values was also leveraged.

Unsupervised Clinical Phenotyping

• Algorithm: Consensus clustering using the K-Means algorithm with Euclidean distance.
• Preprocessing: Continuous variables were Z-score standardized; categorical variables were one-hot encoded.
• Optimization: The optimal number of clusters was determined to be four based on silhouette coefficients and clinical interpretability.
• Validation: Phenotype centroids derived from MIMIC-IV were projected onto the eICU cohort to verify stability.

Predictive Modeling

• Algorithm: Extreme Gradient Boosting (XGBoost) was selected for its ability to handle nonlinear interactions and complex feature spaces.
• Training: 5-fold cross-validation with early stopping (patience = 20) was used on the MIMIC-IV cohort.
• Hyperparameters: Maximum tree depth of 4, learning rate of 0.05, and binary logistic loss.
• Interpretability: The SHapley Additive exPlanations (SHAP) framework was applied to quantify feature contributions and visualize nonlinear risk patterns.

3. Key Results

Clinical Phenotypes

Four distinct, data-driven phenotypes were identified and consistently observed in both cohorts:

1. Phenotype 0 (Low-risk homeostasis): The largest group (50–60%). Characterized by stable physiology, low organ dysfunction scores, and the lowest in-hospital mortality (10%).
2. Phenotypes 1 & 2 (Intermediate-risk progression): Groups with moderate physiological stress and laboratory abnormalities. Mortality rates ranged from ~19% to 22%.
3. Phenotype 3 (High-risk multi-organ failure): The smallest group (9–16%) but the most severe. Characterized by markedly elevated BUN, lactate, and high dependency on mechanical ventilation and vasopressors. Mortality rates were the highest (45–50%).

• Validation: The four-phenotype structure demonstrated robustness across both datasets, with distinct survival trajectories confirmed by Kaplan-Meier analysis (Log-rank P < 0.001).

Predictive Performance

• Discrimination: The XGBoost model achieved an Area Under the Curve (AUC) of 0.818 (95% CI 0.808–0.828) in the derivation cohort and 0.811 (95% CI 0.798–0.824) in the external validation cohort.
• Comparison: The model numerically outperformed traditional severity scores (SOFA AUC 0.769; APACHE IV AUC 0.742) and logistic regression (AUC 0.812).
• Calibration: The model showed acceptable calibration with Brier scores of 0.129 (derivation) and 0.132 (validation).
• Subgroup Analysis: Performance remained generally stable across age, sex, cancer type, and severity strata, though discrimination was slightly attenuated in the highest-acuity subgroups.

Model Interpretation (SHAP)

• Key Drivers: Blood urea nitrogen (BUN), respiratory rate, mechanical ventilation, and vasopressor use were identified as the top contributors to predicted mortality.
• Nonlinear Patterns: SHAP dependence plots revealed nonlinear risk thresholds. For instance, BUN and respiratory rate showed a gradual risk increase at lower values, followed by a sharp rise once specific transition ranges were exceeded (approx. 20–30 mg/dL for BUN and 20–25 breaths/min for respiratory rate).

4. Significance and Claims

The authors claim that this study provides a framework for early risk stratification in critically ill cancer patients using only routinely available data from the first 24 hours of ICU admission. The significance of the work lies in three main areas:

1. Phenotype Discovery: The identification of four clinically distinct phenotypes offers a more granular view of patient heterogeneity than traditional severity scores, potentially aiding in the identification of patients with "futile" trajectories (Phenotype 3) versus those with preserved physiological reserve (Phenotype 0).
2. Explainable AI: By integrating SHAP, the study addresses the "black box" limitation, providing clinicians with patient-level risk attribution and revealing nonlinear physiological patterns that linear models miss.
3. Generalizability: The model's performance was maintained across two distinct healthcare systems (MIMIC-IV and eICU) despite structural differences in diagnostic coding (e.g., lack of specific cancer subtypes in eICU), suggesting the model captures fundamental physiological patterns of deterioration rather than dataset-specific artifacts.

The paper concludes that while the model shows promise as an early risk-stratification tool, it requires prospective evaluation in diverse healthcare systems and clinical implementation studies to confirm its utility in real-world workflows. The authors emphasize that the SHAP-derived transition ranges should be viewed as hypothesis-generating reference points for monitoring rather than definitive treatment thresholds.