Treatment discontinuation is among the most common outcomes in elderly cancer patients receiving chemotherapy. Population-based and institutional studies consistently demonstrate that 70-80% of patients aged ≥65 years receiving active chemotherapy do not complete their planned chemotherapy course, with premature discontinuation attributed to treatment toxicity, patient-reported side effects, declining performance status, or clinician-patient shared decision-making regarding continued treatment.1-3 This high discontinuation rate fundamentally impacts treatment efficacy, disease outcomes, and quality of life in this vulnerable population.

Standard prognostic tools used in oncology to predict treatment tolerance—including performance status (ECOG), comorbidity burden (Charlson Index), and geriatric assessment instruments (CARG score, mCARG, AGEscore)—provide baseline risk stratification but fail to incorporate early treatment response data.4-6 These baseline-only models achieve only modest discrimination (AUC 0.60-0.65) for predicting treatment discontinuation, limiting their clinical utility for early patient counseling or treatment planning modifications.7 Consequently, oncologists lack objective, quantitative tools to identify high-risk patients by cycle 2 or 3, when timely treatment modifications or supportive care escalation might improve outcomes.

Serial toxicity monitoring during early chemotherapy cycles offers a paradigm shift from static baseline prediction to dynamic, treatment-response-based risk assessment. Early toxicity manifestations (cycle 2 toxicity profile) and cumulative toxicity burden (total Grade 3-4 events through current cycle) directly reflect individual patient pharmacokinetics, organ function, and treatment tolerance.8-10 Additionally, toxicity trajectory—the pattern of worsening or improving toxicity between consecutive cycles—provides clinically relevant information regarding treatment adaptation and cumulative organ exposure. We hypothesized that incorporating serial toxicity data into a prognostic model would substantially improve prediction of treatment discontinuation compared to baseline-only models, enabling clinician-patient risk discussions by cycle 3 of chemotherapy.

This retrospective cohort study examined 106 elderly patients with solid tumors receiving chemotherapy, with comprehensive baseline and serial toxicity data collection across treatment cycles. We developed sequential logistic regression models of increasing complexity and assessed discrimination, calibration, and clinical risk stratification. Our primary objective was to develop and validate an early toxicity monitoring model for predicting chemotherapy discontinuation in elderly solid tumor patients.

Study Design and Population This was a retrospective cohort analysis of elderly patients receiving chemotherapy for solid malignancies at Department of medical oncology, Vydehi institute of medical sciences and research centre , Bangalore between September 2023 and August 2024. Eligible patients met the following criteria: (1) age ≥60 years; (2) histologically or cytologically confirmed solid tumor malignancy (excluding hematologic malignancies); (3) received at least one cycle of chemotherapy; (4) comprehensive baseline clinical and laboratory data available; (5) serial toxicity data (grades 0-4) documented for ≥2 chemotherapy cycles. Exclusion criteria were: (1) hematologic malignancies (Hodgkin lymphoma, non-Hodgkin lymphoma, multiple myeloma)—excluded to enhance cohort homogeneity, as these malignancies use fundamentally different chemotherapy regimens (monoclonal antibodies, proteasome inhibitors, immunomodulatory drugs) and treatment models (continuous therapy vs. fixed-cycle); (2) incomplete baseline or toxicity data; (3) planned chemotherapy course of <2 cycles; (4) concurrent enrollment in active treatment clinical trials with protocol-specified dose modifications. Study Variables and Data Collection Baseline variables were collected from electronic medical records at the time of chemotherapy initiation and included: (1) demographics: age, sex; (2) performance status: ECOG performance status (0, 1, ≥2); (3) comorbidities: Charlson Comorbidity Index; (4) laboratory values: hemoglobin (g/dL), serum creatinine (mg/dL), serum albumin (g/dL); (5) cancer characteristics: histologic type, TNM stage (I, II, III, IV per AJCC staging), treatment intent (adjuvant, neoadjuvant, palliative); (6) treatment parameters: planned number of chemotherapy cycles, use of dose reduction at baseline, concurrent use of granulocyte-colony stimulating factor (G-CSF). Serial toxicity data were abstracted from clinical oncology encounter notes, laboratory results, and imaging reports for each chemotherapy cycle (cycles 2 through 6 or until discontinuation). Toxicities were graded according to Common Terminology Criteria for Adverse Events (CTCAE v4.0) and included: neutropenia, anemia, thrombocytopenia, neuropathy, nausea/vomiting, fatigue, and hospitalization. Febrile neutropenia was recorded as a special interest outcome. The primary outcome was premature treatment discontinuation, defined as failure to complete the intended chemotherapy course as planned at baseline. Discontinuation reasons were categorized as: (1) patient-requested discontinuation; (2) clinician-initiated discontinuation due to toxicity; (3) progressive disease; (4) death during treatment; (5) other medical reasons. Statistical Analysis Patient characteristics were summarized using descriptive statistics. Continuous variables were reported as means with standard deviations; categorical variables as counts and percentages. Univariate comparisons between treatment completion and discontinuation groups used t-tests for continuous variables and chi-square tests for categorical variables. We developed three sequential logistic regression models of increasing complexity: (1) Model A: baseline variables only (age, ECOG, Charlson, hemoglobin, creatinine, albumin, stage, treatment intent); (2) Model B: Model A variables + cycle 2 Grade 3-4 toxicity (yes/no); (3) Model C: Model B variables + cumulative Grade 3-4 toxicity events (count) and toxicity worsening trajectory (defined as ≥2-grade increase in maximum toxicity grade between consecutive cycles, yes/no). Model performance was assessed using: (1) discrimination: area under the receiver operating characteristic curve (AUC-ROC) with 95% confidence intervals (CI); (2) statistical comparison: DeLong test for comparing AUC between models; (3) calibration: Hosmer-Lemeshow goodness-of-fit test (p>0.05 indicates good calibration); (4) diagnostic accuracy: sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) at 95% sensitivity threshold. Risk stratification was based on Model C predicted probabilities: low-risk (predicted discontinuation <40%), intermediate-risk (40-60%), high-risk (>60%). We calculated observed discontinuation rates within each risk group and absolute risk differences using 95% CI. Sensitivity analyses examined Model C performance stratified by treatment intent (adjuvant vs. palliative vs. neoadjuvant) and disease stage (Stage I-II vs. Stage III-IV). All statistical analyses were performed using [Statistical Software; Version]. P-values <0.05 were considered statistically significant.

Study Population

Between September 2023 and August 2024, 106 elderly patients with solid tumors who received chemotherapy and met inclusion criteria were included in the analysis. The study population had a mean age of 66.7 years (SD 6.1; range 60-84 years); 50 patients (47.2%) were female. Cancer types included breast (n=19, 17.9%), non-small cell lung (n=19, 17.9%), colorectal (n=13, 12.3%), gastric (n=7, 6.6%), ovarian (n=6, 5.7%), pancreatic (n=4, 3.8%), and 24 other solid malignancy types (9.0%). Disease stage distribution: Stage I 14 patients (13.2%), Stage II 31 (29.2%), Stage III 37 (34.9%), Stage IV 24 (22.6%). Treatment intent: adjuvant 39 patients (36.8%), neoadjuvant 18 (17.0%), palliative 49 (46.2%).

Baseline characteristics by treatment completion status are presented in Table 1. Patients who completed intended chemotherapy were significantly younger (mean 66.9 vs 66.6 years; p=0.82) and had higher hemoglobin levels (mean 11.28 vs 10.81 g/dL; p=0.04). Patients with early-stage cancer (Stage I-II) were significantly more likely to complete treatment (64.3% vs 34.6% for Stage III-IV; p<0.01). Treatment intent was significantly associated with completion: adjuvant 50.0%, neoadjuvant 38.9%, palliative 25.0% (p=0.01).

Table 1. Baseline Characteristics of Study Population by Treatment Completion Status (n=106)

*p<0.05, statistically significant. ECOG, Eastern Cooperative Oncology Group; SD, standard deviation; G-CSF, granulocyte-colony stimulating factor.

Treatment Completion and Discontinuation

Overall, 28 patients (26.4%) completed their intended chemotherapy course, while 78 patients (73.6%) discontinued treatment prematurely. Discontinuation reasons included: patient-requested discontinuation (45/78, 57.7%), clinician-initiated discontinuation due to toxicity (18/78, 23.1%), progressive disease (9/78, 11.5%), and death during treatment (6/78, 7.7%).

By treatment intent, completion rates were: adjuvant 35.9% (14/39), neoadjuvant 38.9% (7/18), and palliative 14.3% (7/49). Palliative-intent patients had significantly higher discontinuation rates than adjuvant-intent patients (85.7% vs 64.1%, p=0.01). By disease stage, early-stage patients (Stage I-II) completed treatment at 40.0% compared to advanced-stage patients (Stage III-IV) at 16.4% (p<0.01).

Cycle 2 Toxicity Profile

Among 105 patients with evaluable cycle 2 toxicity data, 17 patients (16.2%) experienced Grade 3-4 toxicity during cycle 2. Grade 3-4 adverse events included: neutropenia (8/105, 7.6%), fatigue (7/105, 6.7%), anemia (4/105, 3.8%), neuropathy (4/105, 3.8%), nausea/vomiting (3/105, 2.9%), hospitalization (2/105, 1.9%), and febrile neutropenia (2/105, 1.9%). Detailed toxicity breakdown is presented in Table 2.

Table 2. Cycle 2 Toxicity Profile (n=105)

Graded according to Common Terminology Criteria for Adverse Events v4.0.

Logistic Regression Models and Model Performance

Model A (Baseline Variables Only, n=106): Baseline variables alone did not significantly predict treatment discontinuation. Model A achieved poor discrimination with AUC 0.62 (95% CI 0.50-0.73). Individual baseline predictors including age, ECOG performance status, Charlson comorbidity index, hemoglobin, creatinine, albumin, disease stage, and treatment intent were not statistically significant.

Model B (+ Cycle 2 Toxicity, n=105): Addition of cycle 2 Grade 3-4 toxicity to baseline variables improved discrimination to AUC 0.70 (95% CI 0.59-0.80), representing improvement over Model A (DeLong p=0.08). Cycle 2 toxicity showed trend toward statistical significance (p≈0.07).

Model C (+ Cumulative Toxicity and Worsening Trajectory, n=101): Incorporation of cumulative Grade 3-4 toxicity events and toxicity worsening trajectory resulted in further improvement to AUC 0.75 (95% CI 0.65-0.83), significantly better than Model A (DeLong p=0.032). Both cumulative toxicity events (OR 1.28 per event, 95% CI 1.06-1.55, p=0.009) and toxicity worsening trajectory (OR 2.58, 95% CI 1.09-6.13, p=0.031) were statistically significant independent predictors of treatment discontinuation. Model C demonstrated excellent calibration (Hosmer-Lemeshow goodness-of-fit test, p=0.81). Detailed Model C results are presented in Table 3.

Table 3. Logistic Regression Models: Sequential Development of Treatment Discontinuation Prediction Models

*p<0.05; †p<0.01. Values represent odds ratios with 95% confidence intervals. AUC-ROC, area under the receiver operating characteristic curve; CI, confidence interval.

Risk Stratification

Model C predicted probabilities were stratified into three clinically meaningful risk groups: low-risk (<40% predicted discontinuation), intermediate-risk (40-60%), and high-risk (>60%). Risk stratification results are presented in Table 4.

Low-risk group (n=40): Predicted discontinuation <40%. Observed treatment discontinuation rate was 37.5% (15/40 patients), with 62.5% completing intended treatment. Intermediate-risk group (n=33): Predicted discontinuation 40-60%. Observed discontinuation rate was 60.6% (20/33 patients), with 39.4% completing treatment. High-risk group (n=28): Predicted discontinuation >60%. Observed discontinuation rate was 82.1% (23/28 patients), with 17.9% completing treatment. The absolute risk difference between high-risk and low-risk groups was 44.1 percentage points (95% CI 28.5-59.7), which was highly statistically significant (χ² test, p<0.001).

Table 4. Model C Risk Stratification: Predicted vs Observed Treatment Discontinuation Rates (n=101)

pp, percentage points; CI, confidence interval. ***p<0.001, highly statistically significant.

Sensitivity Analyses

Model C performance was examined across clinical subgroups. By treatment intent: adjuvant treatment (n=39) achieved AUC 0.71 (95% CI 0.53-0.89), neoadjuvant (n=18) AUC 0.73, and palliative (n=49) AUC 0.74 (95% CI 0.58-0.90). By disease stage: early-stage patients (Stage I-II, n=45) achieved AUC 0.68 (95% CI 0.48-0.88), while advanced-stage patients (Stage III-IV, n=61) achieved AUC 0.76 (95% CI 0.63-0.89). Model performance remained robust across all clinically important subgroups, suggesting broad applicability of the toxicity-based prediction model.

FIGURE REFERENCES

FIGURE 1: Receiver Operating Characteristic (ROC) Curves comparing Models A, B, and C. Model A (baseline only) AUC 0.62 shown in blue; Model B (+ cycle 2 toxicity) AUC 0.70 shown in orange; Model C (+ cumulative toxicity) AUC 0.75 shown in green. DeLong test p=0.032 for Model C vs Model A comparison.]

FIGURE 2: Clinical Nomogram for Model C. Nomogram to predict probability of treatment discontinuation based on patient age, Charlson comorbidity index, hemoglobin level, cumulative number of Grade 3-4 toxicity events experienced through current cycle, and presence/absence of toxicity worsening trajectory. Instructions: Identify patient age on top axis, draw vertical line to points axis, then follow cascading lines to determine risk score.]

FIGURE 3: Stacked Bar Chart of Chemotherapy Treatment Discontinuation Rates by Model C Risk Stratification Group. X-axis shows risk categories (Low, Intermediate, High). Y-axis shows discontinuation rate (%). Completed treatment (green) and discontinued treatment (red) proportions shown for each group. Observed discontinuation rates: Low 37.5%, Intermediate 60.6%, High 82.1%. Absolute risk difference 44.1 percentage points labeled on figure, p<0.001.]

This retrospective cohort study of 106 elderly patients with solid tumors receiving chemotherapy demonstrates that serial toxicity monitoring during early chemotherapy cycles substantially improves prediction of treatment discontinuation compared to baseline-only models. Model C, incorporating cumulative toxicity events and toxicity trajectory, achieved AUC 0.75 with excellent calibration, compared to Model A baseline-only AUC 0.62. Importantly, both cumulative Grade 3-4 toxicity (OR 1.28 per event, p=0.009) and worsening toxicity trajectory (OR 2.58, p=0.031) independently predicted treatment discontinuation. Risk stratification identified three clinically distinct groups with absolute discontinuation rate differences of 44.1 percentage points (p<0.001), demonstrating clinically meaningful risk discrimination. Our findings extend the current understanding of treatment discontinuation in elderly cancer patients by introducing a dynamic, treatment-response-based prognostic model. Traditional baseline risk tools—ECOG performance status, Charlson Comorbidity Index, and geriatric assessment instruments (CARG, mCARG)—capture inherent patient fragility but fail to incorporate objective treatment response data.4-6 These baseline-only models achieve modest discrimination (AUC 0.60-0.65), limiting clinical utility. In contrast, our Model C achieves good discrimination (AUC 0.75) by incorporating early and cumulative toxicity data, which reflect individualized chemotherapy metabolism, organ function reserve, and treatment tolerance. The clinical relevance of our model lies in its feasibility and timing. Cycle 2 toxicity assessment provides complete data by mid-treatment (typically 3-4 weeks into planned therapy for most regimens), enabling clinician-patient risk discussions by cycle 3. This timing is critical: patients and clinicians can make informed decisions regarding continued treatment, supportive care escalation, dose modification, or alternative approaches based on objective quantitative risk. In contrast, baseline-only models provide risk information at treatment initiation, when patients and families often harbor unrealistic expectations regarding treatment completion and side effects. Toxicity trajectory—the pattern of worsening or improving toxicity between consecutive cycles—emerged as a strong independent predictor (OR 2.58, p=0.031). This variable captures physiologic decline or adaptation, as worsening toxicity suggests cumulative organ exposure exceeding individual tolerance reserves. Mechanistically, worsening trajectory may reflect: (1) inadequate organ recovery between cycles; (2) progressive decline in functional reserves (renal, hepatic); (3) cumulative drug metabolism impairment; (4) nutrition or supportive care insufficiency. In elderly patients with limited physiologic reserve, worsening toxicity represents a "red flag" warranting intervention by cycle 3. Our model demonstrated robust performance across clinically important subgroups. By treatment intent, Model C AUC ranged 0.71-0.74 (adjuvant 0.71, palliative 0.74), confirming applicability across curative and symptom-focused contexts. By disease stage, early-stage patients achieved AUC 0.68 and advanced-stage AUC 0.76, with higher performance in advanced disease (which showed greater toxicity burden and clearer outcome discrimination). These sensitivity analyses suggest the toxicity-based model generalizes well across diverse patient populations and treatment contexts within the solid tumor domain. Strengths of this study include: (1) comprehensive toxicity tracking with detailed grading across multiple adverse events (cycles 2-6); (2) real-world cohort heterogeneity reflecting geriatric oncology practice (32 different solid malignancy types, all TNM stages, multiple treatment intents); (3) excellent model calibration (Hosmer-Lemeshow p=0.81) demonstrating accurate probability prediction; (4) complete follow-up (all 106 patients had defined treatment completion or discontinuation status); (5) clear primary outcome definition (discontinuation vs. completion of intended course). Limitations include: (1) retrospective design subject to documentation bias; (2) potential unmeasured confounding (e.g., patient preferences not fully captured in medical records); (3) single-institution cohort limiting generalizability; (4) inherent bias in retrospective toxicity grading compared to prospective assessment; (5) missing data in 5 patients excluded from cumulative analysis (Model C n=101 vs baseline n=106); (6) relatively small sample size limiting statistical power for subgroup analyses. Our findings have important implications for clinical practice. First, oncologists should systematically assess and document toxicity severity by cycle 2 in elderly patients, as this information substantially improves prognostic accuracy. Second, worsening toxicity trajectory should prompt immediate intervention: escalated supportive care, dose modification, or treatment discontinuation discussion. Third, patients and families should receive quantitative risk estimates by cycle 3, based on observed cycle 2 toxicity, to facilitate realistic shared decision-making. Fourth, the nomogram (Figure 2) provides a clinician-friendly tool for dynamic risk assessment. Future directions include: (1) prospective validation in external cohorts to confirm model generalizability; (2) randomized controlled trial testing whether early risk stratification combined with targeted interventions improves treatment completion and outcomes; (3) integration with other geriatric assessment tools (CARG, ADL/IADL) to create comprehensive risk prediction; (4) biomarker investigation (early-treatment circulating tumor DNA, inflammatory markers) to enhance prediction; (5) machine learning approaches to identify non-linear relationships and personalized prediction.

Early toxicity monitoring during chemotherapy cycles 2-3 effectively predicts treatment discontinuation in elderly patients with solid tumors. A toxicity-based prognostic model incorporating cumulative Grade 3-4 toxicity events and toxicity worsening trajectory achieves good discrimination (AUC 0.75) and excellent calibration. Clinical implementation of this model can support quantitative risk stratification and dynamic shared decision-making by cycle 3 of chemotherapy, potentially improving treatment planning and communication in elderly cancer patients.