We performed a prospective multi-center cohort study in solid cancer patients scheduled to receive active treatment in medical oncology departments of three medical centers: First Affiliated Hospital of Zhengzhou University, Sun Yat-Sen Cancer Center, and Affiliated Cancer Hospital of Shantou University Medical College. Recruitment for the current study started in 2020.1 and ended in 2021.8 and all patients were consecutively recruited (See Participant Eligibility Criteria). The primary aim was to compare objective and subjective CRCD in patients receiving chemotherapy alone (chemo group) and ICI combined with chemotherapy (chemoICI group).

The research was approved by institutional review boards of the Affiliated Cancer Hospital of Shantou University Medical College (Approval No. ST-ZLY-2020-716). The procedure was performed according to the Helsinki Declaration. Included patients had given written informed consent before participation. Reporting adheres to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) checklist for cohort studies (https://www.equator-network.org/reporting-guidelines/strobe/).

Inclusion criteria of the current study included 1) patients with stage I to IV solid cancers scheduled with standard dose of chemotherapy (chemo) or chemotherapy + ICI at the research sites during study period; 2) age > 30 years; 3) immunotherapy-naive before inclusion for both groups; 3) prognostic survival > 1 year; 4) normal liver function tests and kidney function tests at cancer diagnosis.

Exclusion criteria at recruitment were: 1) known history of CNS tumors, trauma, or metastasis; 2) history of neurodegenerative disease (e.g. Parkinson’s Disease); 3) history of or concurrent targeted therapies of tyrosine kinase inhibitors; 4) concurrent radiotherapy; 5) history of the psychiatric disease; 6) alcohol use disorder or current prescribed/unprescribed narcotic users; 7) brain radiotherapy history; 8) high-risk of metabolic encephalopathy during treatment as evaluated by oncologists.

Participants underwent detailed neurological exams. If any neurological deficits were identified, radiological procedures of MRI were ordered. If brain lesions, including metastasis or other tumors, were found and diagnosed, further follow-up of neurological exams would be excluded. Key exclusion criteria during follow-up included: 1) inter-group change of therapy regimens into chemo or chemoICI group; 2) treatment discontinuation of over 3 months; 3) any medical condition not negotiable to follow-up assessment, including stroke and prolonged drug intoxication; 4) brain infections or metabolic encephalopathy as diagnosed by oncologists.

The baseline assessment was conducted after the treatment modality was determined by oncologists and before treatment began. The cognitive assessment consisted of subjective questionnaires and objective tests, which were measured by Perceived Cognitive Impairment (PCI, the cognitive impairment domain of FACT-cog) and Montreal Cognitive Assessment (MoCA, a screening cognitive test), respectively. The cognitive assessment in this study utilized screening tools rather than comprehensive neuropsychological testing. The MoCA was selected as an objective screening measure due to its sensitivity in detecting mild cognitive impairment, though it does not replace a full neuropsychological battery. While full neuropsychological batteries provide a more comprehensive assessment, they are often impractical in real-world oncology settings due to time and resource constraints. Recent studies have confirmed that MoCA is more effective in assessing executive function, attention, and memory, cognitive domains frequently affected in cancer patients (22, 23).

The PCI scale, a component of the FACT-cog, was included to capture subjective cognitive complaints, which are frequently reported in cancer patients and correlate with quality of life (24). This dual approach aligns with current recommendations for preliminary cognitive screening in oncology settings, providing an accessible method for tracking cognitive changes over time. Practice effects were mitigated by analyzing score change rather than absolute values, though residual learning effects cannot be fully excluded. Subjective measures were administered to participants at recruitment (baseline, time 0) and 4 sessions during follow-up with a 3-month interval: 3, 6, 9, and 12-month sessions during the medical check-up clinic of the research setting, during which time patients had cancer checkups. Objective assessment was carried out at recruitment (baseline, time 0) and 3, 6, and 12-month sessions. Follow-up was set conveniently with medical checkups to minimize dropouts. Patients were given 100 yuan as compensation if all sessions were completed. During each session, blood tests, questionnaires, and psychometric scales were drawn and reported otherwise. Final data were compared between comparable groups in 2022.9 utilizing propensity score matching (PSM). Measure instruments and matching protocols were shown in Supplementary Methods .

Cognitive variables included PCI scores (self-reported cognitive complaints) and MoCA scores (objective cognitive function). PCI assessed memory, attention, and executive function complaints, while MoCA measured global cognition, including memory, executive function, and visuospatial abilities. The analysis focused on score changes over time and the occurrence of meaningful cognitive impairment events, defined by the minimal clinically important difference (MCID) threshold. The primary endpoint was the MoCA and PCI score change from baseline scores. The 0.5*standard deviation (SD) of baseline MoCA and PCI scores after PSM was set as the threshold for MCID, based upon the distribution method of calculus (25). This approach provides a clinically interpretable threshold, and it was widely accepted for preliminary studies in oncology populations (25). Therefore, any score decrease from baseline at any follow-up session, of > MCID would be marked as a meaningful subjective impairment event (MSIE, for PCI) or a meaningful objective impairment event (MOIE, for MoCA). Thus, score change from baseline as well as proportions of MOIE/MSIE could be compared in the chemo versus chemoICI group, and the proportions included prevalence (all cases) and incidence (new cases). As an exploratory outcome, post hoc analysis was done in the chemoICI group to identify the independent association between any incident irAEs and MOIE/MSIE incidence. The irAE was diagnosed as solicited during follow-up according to the National Cancer Institute Common Terminology Criteria for Adverse Events (26).

The only exposure was ICI treatment in primary outcome analysis, with other information set as the confounding variables in PSM analysis or linear regression models (see Supplementary Methods ). Specifically, the demographic variables included age, sex, body mass index values (BMI), illiteracy state, and socioeconomic status (SES). The primary confounders included in the analysis were age, sex, baseline cognitive scores, chemotherapy regimen, and comorbidities. Although education level, occupational status, and premorbid intelligence can influence cognitive function, they were not included due to incomplete data, prioritization of treatment-related factors, and potential recall bias in self-reported premorbid intelligence. Instead, baseline cognitive scores were used as a proxy for pre-existing cognitive function (4). Clinical variables included cancer stage, pathological diagnosis, chemotherapy regimens (platinum-based or others), history of chemotherapy, diabetes diagnosis, and Eastern Cooperative Oncology Group-Performance Scores (ECOG-PS). PCI and MoCA scores, depression status, and opioid use during follow-up were also recorded as baseline variables that might affect cognitive functions (27). To control for learning effects, score change rather than absolute scores were represented and compared in two groups. We assessed quality of life levels using Functional Assessment of Cancer Treatment-General (FACT-G) at baseline (28, 29).

To estimate the sample size needed for statistical significance, we presumed a clinically relevant difference in score change of 4 points between the two groups, as reported in prior observation cohorts (30), and it was calculated > 272 patient pairs to detect such presumed difference, assuming drop-out rate of 20%, with a statistical power of 90% and type I error of 0.05 (31–33).

Propensity scores were calculated with multiple logistic regression and associated variables included all variables as adjusting confounders. In propensity matching, a nearest greedy algorithm was adopted to give head-to-head (1:1) matching between each patient in the two groups. Caliper width of 0.2 was adopted in score matching without replacement. Matching competence was evaluated with standardized mean difference (SMD), but traditional methods of paired tests were also represented for each baseline variable (34). An SMD > (√ ((n1 +n2)/n1*n2))*1.96 was defined as imbalanced matching (n1 = n2 refers to pre-matched sample sizes) (35).

Equivalent statistical tests of difference of score change difference in propensity-score matched samples were performed with Wilcoxon signed-rank tests. Equivalence tests of paired proportion difference, or MOIE/MSIE difference, were performed with McNemar tests. PCI or MoCA score change difference in subgroups with or without incident irAE was compared by linear regression models to adjust for all baseline variables. To calculate incidence rate, or new case rate, of MOIE/MSIE during follow-up, Kaplan-Meier survival curve was adopted to estimate mean event-free survival (EFS) time of comparable groups or subgroups. As for the consideration of sensitivity and specificity, the MoCA has demonstrated sensitivity (82%) and specificity (76%) in detecting mild cognitive impairment in cancer cohorts (22), though its ability to distinguish immune-mediated neurotoxicity remains unvalidated. Misclassification risks due to overlapping irAE symptoms (e.g., fatigue, depression) were mitigated by excluding patients with metabolic encephalopathy or infections during follow-up. Additionally, the PCI scale, while sensitive to patient-reported cognitive complaints, may be influenced by emotional states such as anxiety or depression, which are common in cancer patients undergoing treatment (36). To address this, we adjusted for depression status in our analyses and excluded patients with psychiatric diagnoses. Paired and independent log-rank tests were adopted to compare difference of EFS rate. the power remains over 90% to calculate difference of EFS rate of over 10% difference in both paired and independent log-rank test, assuming two-sided, 5% type I error. Multivariate EFS analysis using proportional hazards model assessed the hazard ratio (95% confidence interval, CI) of irAE that adjusted for all baseline variables, and the continuous variable were not categorized to preserve integrity. Sample sizes were calculated with PASS (version15.0.), and all statistical analysis was performed in R (version 4.0.5) software. Plots of prevalence were drawn with Graphpad Prism (version 8).

At baseline 1557 patients were enrolled and consented to the study protocol ( Figure 1 ), reaching an initial sample size of 1304 (794 in the chemo group and 510 in the chemoICI group). All participants were from the Han ethnic group, and participants with illiteracy (n = 394, 30.2%) had significantly lower baseline PCI scores, but not MoCA scores, than non-illiterate participants. All baseline variables varied significantly between the chemo and chemoICI groups, except for sex, SES, and MoCA scores ( Table 1 ).

Propensity scores were calculated, and post hoc randomization was carried out to 1:1 match 460 pairs of patients from the chemo and chemoICI group, respectively, that dropped unmatched 50 (9.8%) and 334 (42.1%) patients from chemoICI and chemo group, respectively. Our a priori-defined imbalance threshold for any variable was 0.13 (see calculation formula in Supplementary Methods ), and all variables had standardized mean difference (SMD) less than the threshold ( Table 1 ). Non-parametric tests showed all variables were well balanced (p > 0.05) except for cancer diagnosis (p = 0.01). The results showed relatively satisfactory balancing of unmatched variables before PSM. After matching, both groups were similar in terms of baseline characteristics, especially PCI and MoCA scores.

Then, the threshold for MOIE and MSIE was calculated (0.5* SD of baseline scores). Our previously defined MCID of MoCA was -1.98 for both groups. MCID of PCI was -5.05 for the chemoICI group and -5.22 for the chemo group. Using this threshold, the statistical power of difference (< -5 or < -2) was calculated to be > 90% in the paired samples assuming a two-sided, type I error of 0.05. In addition, ICI types were relatively evenly distributed ( Supplementary Table 1 ) in the matched chemoICI group.

Chemotherapy-only (chemo) group was over-sampled to allow for maximum matching for the chemotherapy plus immune checkpoint inhibitor (chemoICI) group. In total, 115 (18.4%) patients in the chemoICI group dropped or were excluded, and 138 (14.8%) patients dropped or were excluded from the chemo group, and the demographics of the excluded patients, including age and sex, did not differ from that of the included patients. All participants were of the Han ethnic group. The reasons for dropout and exclusion were in Supplementary Materials.

MoCA score change were significantly different between chemoICI and chemo group in the 12-month session (p < 0.01), but not in the 3-month (p = 0.96) or 6-month (p = 0.55) session ( Figure 2A ). Within the chemoICI group, adjacent score change were all significantly different from baseline to 12 months (p < 0.05, p < 0.05, and p < 0.001), suggesting progressive decline ( Figure 2B ).

PCI score change were significantly different between the chemoICI and chemo group in the 6-month (p < 0.05), 9-month (p < 0.001), and 12-month (p < 0.001) sessions, but not in the 3-month (p = 0.36) session ( Figure 2C ). Within the chemoICI group, adjacent score change were significantly different from baseline to 9-month (p < 0.001, p < 0.05, and p < 0.001) session but not in the 12-month (p = 0.13) session ( Figure 2D ).

Prevalence of MOIE was growing with time in both groups but was significantly different only in the 12-month session (40.0% in chemoICI vs. 26.5% in chemo, Figure 3A ) during follow-up. The twelve-month incidence rate of MOIE was 0.35 (chemo) vs. 0.44 (chemoICI). Kaplan-Meier survival curve estimated EFS time of chemoICI to be 10.88 (95%CI 10.63 - 11.13) months, significantly shorter (p < 0.01) than EFS of chemo (mean = 11.15, 95%CI 10.93 - 11.37 months, Figure 3B ).

The prevalence of MSIE was also growing with time in both groups and was significantly different in all sessions ( Figure 3C ). Kaplan-Meier survival curve estimated EFS time of chemo was estimated to be significantly shorter than chemoICI (p < 0.01, Figure 3D ), with MSIE incidence rate of 0.56 in chemoICI and 0.38 in chemo group.

An exploratory, post hoc analysis was done to identify the independent relationship between irAE and cognitive impairment in the chemoICI group. Overall, 143 of 460 patients (31.1%) reported and were diagnosed with irAEs. After adjusting for all baseline variables, irAE was significantly associated with the 12-month incidence of MSIE and MOIE in the chemoICI group (p < 0.01 for both, Supplementary Table 2 ). The hazard ratio was 2.2 (95%CI 1.64 - 2.94) for MOIE and 2.11 (95%CI 1.59 - 2.8) for MSIE, and there was a significant difference of mean EFS time in patients with and without irAE ( Figures 4A, B ). Significantly higher MCOA and PCI score change were found in patients with irAEs ( Figures 4C, D ) that are adjusted for all baseline variables.