Up to 75% of survivors of cancers that originate outside of the central nervous system demonstrate cognitive impairment before, during or following cancer treatments (1–4). CRCI typically presents as difficulties with attention, executive function, memory, and processing speed that can last months or years after the conclusion of treatment (1–5). CRCI can have debilitating effects on survivors’ quality of life including their social and occupational functioning (3, 4, 6–9), and is associated with increased death including cancer-related and all-cause mortality (10–12). Millions of cancer survivors [16.9 million in 2019 alone (13)] are at risk for CRCI.

Cognitive symptoms vary greatly from person to person in terms of domains impacted, as well as symptom severity and trajectory (e.g., late onset, persistent, improving over time). In fact, most symptoms associated with treatment toxicities or cancer pathology are highly variable and unique to each survivor. Thus, recent calls to action have been made for precision survivorship in the field of oncology, where some genetic risks for treatment-related toxicities have been identified in pediatric oncology populations (14). Since 2016, the National Cancer Institute (NCI), has prioritized precision medicine which includes “precision oncology” and “precision survivorship” (15).

The NCI defines precision medicine as using a person’s own biology to diagnose cancer or determine a course of treatment for cancer. Many advancements have been made in precision medicine with respect to cancer diagnostics and treatments, but there has been a lag in applying precision medicine to cancer survivorship, especially for adult cancer survivors (16). Pre-treatment data could be used to determine who is most susceptible to treatment related toxicities, including CRCI.

Precision medicine approaches to CRCI may significantly improve the diagnosis and prognosis of this syndrome. Under or over diagnosing of CRCI incurs significant costs such as negative individual, familial (including impact to the caregiver), and societal burden. The economic benefits of accurate and timely diagnosis have been shown to significantly reduce health care costs. For example, one study showed that health care costs would be reduced by more than half to two-thirds if a diagnosis had been made earlier (17). Patients with CRCI often experience increased stress, anxiety, and depression, loss of employment opportunities, and more social isolation. Patients’ distress can be significantly reduced by health-related perceptions of situational control, validation, and course of deficit, all of which are impossible without a correct diagnosis. Understanding and correctly characterizing the CRCI condition is also imperative for precisely developing and guiding patients toward interventions that can ameliorate their symptoms.

Clinical symptoms of CRCI are currently assessed using standardized neuropsychological tests and/or self-report measures. Guidelines for using neuropsychological tests in CRCI research were published in 2011 (18). These tests provide insight regarding an individual’s cognitive strengths and weaknesses and can assist with diagnosing neuropsychiatric conditions. However, several reports have suggested that these tests may have limited sensitivity, specificity, and reliability for CRCI (19–22). Many of these measures were developed to assess severe neuropathology (21, 23) and therefore may have ceiling effects for CRCI. These tests tend to lack ecological validity (24, 25) and there is also evidence of ethnic or racial biases in standardized neuropsychological testing (26–28).

Guidelines for using self-report measures in CRCI research were published in 2021 (29). Self-report measures reflect daily functioning (30), correlate with mood and psychological factors (24, 30–32), and some were developed specifically for cancer survivor populations (33, 34). Incidences of CRCI are higher when subjective measures are used (3) suggesting they may be more sensitive to this syndrome. However, self-report measures also have several limitations including susceptibility to retrospective recall, response, and state-dependent biases as well as demand characteristics (35–38). Importantly, some studies have found elevated response bias among cancer survivors compared to normative groups (36, 37, 39). Self-reported CRCI has been criticized for rarely correlating with neuropsychological performance (31). However, it is likely that self-report and neuropsychological tests measure different aspects of CRCI and cannot be considered interchangeable (40). Accordingly, evidence suggests that subjective and objective measures of CRCI represent different neural phenotypes. Both objective and subjective measures of CRCI have been shown to correlate with structural and functional brain abnormalities, but these appear to be largely distinct and non-overlapping (41–45).

Neither standardized cognitive tests nor self-report measures can track an individual’s day-to-day variability in cognitive function. Both rely on population-based methods (between group variability) rather than precision-based methods (within person variability). Cognitive functioning is subject to individual variability and real-life demands, thus using one’s own cognitive variability within one’s own environment may be a more ecologically valid and unbiased way of assessing CRCI. A precision health approach to defining CRCI, which is an individually focused approach, is especially beneficial for historically marginalized groups, such as racial/ethnic minorities, as mentioned above.

One of the greatest challenges to clinical management of CRCI is its inconsistent definition within the existing literature. As noted above, CRCI is currently evaluated using neuropsychological test performance. However, these tests rarely provide normative cutoff scores for deficient performance. Therefore, classification of impairment based on these tests in patients with cancer has relied on arbitrary cutoff scores leading to inconsistent results (46). To better harmonize CRCI studies, the International Cognition and Cancer Task Force (ICCTF) suggested a specific metric for determining impairment based on z-scores (18). Few studies have compared the ICCTF definition to other rubrics, although one study showed it was nearly twice as sensitive to impairment compared to other definitions (47). However, different results occur depending on the reference group used for z-score calculation and the size and composition of the testing battery (48, 49).

A major limitation of the z-score and similar approaches, including standardized based regression (50) and reliable change index (22), is that they result in dichotomous classification (yes/not impaired; yes/no declined). Dichotomous classification is not consistent with the complex, continuous nature of brain function. Akin to “black and white thinking,” dichotomization ignores dimensional nuances that could provide greater insight regarding cognitive function (51). In fact, a recent study demonstrated that binary impairment definitions, including the ICCTF criterion, yield unreliable cognitive classifications (52). This is particularly important for CRCI given the historical controversy regarding its existence. Specifically, many patients are classified as “unimpaired” on neuropsychological tests despite experiencing significant difficulties completing everyday cognitive tasks. This inconsistency may suggest that there are latent subgroups of patients whose cognitive changes do not fit into black and white categories.

Several statistical techniques exist that can determine subgroups within the data that are not readily apparent (Table 1). For example, growth mixture modeling, latent profile analyses and clustering have been used to discover CRCI subgroups in a data-driven manner. Growth mixture modeling determines latent subgroups based on different trajectories of an outcome across longitudinal timepoints (69). Prior studies of CRCI using growth mixture modeling have identified up to five latent subgroups of self-reported or objective cognitive function, with the most common finding being three subclasses (53–58).

Once the latent subgroups are identified, characteristics of the subgroups can be examined to determine their clinical significance including factors associated with risk and resilience. For example, using growth mixture modeling, Merriman et al. (54) identified three latent subgroups of subjective CRCI in patients with breast cancer. They labeled the subgroups “more frequent,” “persistent,” and “almost never” based on frequency of cognitive complaints. They found that patients in the “more frequent” subgroup had unique profiles of treatment regimens, genotypes, and psychiatric symptoms compared to the other subgroups (54). Alternatively, covariates can be included in the initial model to examine their impact on the longitudinal outcome (70).

Latent profile analysis also seeks to identify unknown subgroups within a larger population using a probabilistic model that describes the distribution of the data (71). Latent profile analysis is appropriate for cross-sectional and longitudinal data. Like growth mixture modeling, latent profile analysis allows for inclusion of covariates to predict latent subgroup membership. Latent profile analysis has been applied to study CRCI subgroups in both pediatric and adult cancers (59–64).

Other latent general linear model analysis techniques have been used to provide insight regarding the association of self-reported cognitive function and neuropsychological tests in patients with breast cancer (65). However, real-world data do not always meet linear model assumptions. Clustering is another statistical technique useful for discovering latent subgroups which is appropriate for linear or nonlinear cross-sectional or longitudinal data. Clustering is based on distances between data points and aims to identify subgroups within a sample that have both high within-group similarity and high between-group dissimilarity. Very few studies have applied clustering to examine profiles of CRCI. Romero and colleagues identified four subtypes of neuropsychological testing scores using k-means clustering (66). The authors interpreted these findings at the test level in terms of which tests were sensitive to impairment, but their results suggested subgroups of participants with different levels of impairment (66). Another study employed k-means clustering as a sensitivity analysis for their latent profile analysis, both of which identified only a dichotomous solution (59).

These studies represent advances in the literature towards a more precise definition of CRCI. However, the mechanisms that result in various symptoms must be known to develop treatments, particularly the mechanisms that differentiate patients who respond better to one treatment versus another. Mechanistic knowledge cannot be derived from symptom assessments alone given that similar symptoms can result from different pathologies (72). Neuropsychological tests, for example, can identify memory, attention, or other cognitive domain deficits. However, these tests summarize multiple cognitive processes into one score and thus are too broad for identifying the precise neural mechanisms involved. Novel data driven approaches as used in the studies above are more granular and able to classify patients into different subtypes based on behavioral data, but the resultant subtypes do not consider the differences in pathophysiology that would dictate precision treatment.

The biologic mechanisms of cognitive disorders are frequently examined using neuroimaging techniques. Multiple magnetic resonance imaging (MRI) studies have demonstrated that CRCI is a brain-based syndrome characterized by significant changes in brain function and structure. Breast cancer survivors who receive chemotherapy treatment tend to show more pronounced global, cortical gray matter and subcortical white matter volume loss when compared with chemotherapy naïve and noncancer controls [see reviews by Sousa et al. (41), Schroyen et al. (42), and Niu et al. (73)]. McDonald et al. (74) reported the first prospective, longitudinal volumetric MRI study of CRCI. They observed a significant reduction over time in regional gray matter density from pre-chemotherapy to 1-month post-chemotherapy followed by partial improvement over time (74). Gray matter alterations have also been demonstrated in very long-term survivors (75–77). We uniquely used volumetric MRI to estimate cortical brain age and demonstrated that it is significantly increased from pre- to post- breast cancer chemotherapy (78). We also distinctively demonstrated that chemotherapy treated breast cancer survivors are significantly more likely to be classified as having incipient Alzheimer’s disease based on volumetric brain network organization compared to chemotherapy naïve survivors (79).

Pathologic white matter changes determined using T2-weighted MRI in patients with cancer exposed to chemotherapy were first described in the nineties (80–82). These changes tended to occur acutely after chemotherapy completion but persisted at 1 year of follow-up (81). Diffusion tensor imaging (DTI) has demonstrated abnormal white matter integrity in the corpus callosum of chemotherapy treated patients with breast cancer compared to noncancer controls (83). Subsequent studies have shown diffuse white matter integrity abnormalities that correlate with cognitive impairment (84, 85), indicating that a demyelinating process may partly underlie CRCI (75). DTI was also uniquely used by our group to computationally simulate the effects of aging on white matter organization, demonstrating that chemotherapy treated breast cancer survivors have lower resilience to brain aging compared to non-cancer controls (86).

Task-based functional MRI (fMRI) studies suggest that certain cognitive tasks are more challenging for patients with breast cancer as they demonstrate expanded recruitment of brain regions to maintain task accuracy. Several lines of research suggest that chemotherapy used in treatment of breast cancer upregulates neural activity (82, 87–91). Brain hypoactivation compared to non-cancer controls has also been reported in several cross-sectional studies (92, 93), yet brain hyperactivation is more often evident over time and is associated with self-reported cognitive complaints (90, 91, 94, 95).

Hyperactivation may therefore help explain the discrepancy that is commonly observed between neuropsychological tests and self-reports of cognitive function. Specifically, patients with breast cancer often demonstrate normal objective task performance but report significant cognitive difficulties (21, 96). Hyperactivation is believed to reflect neural compensation in response to brain injury, which can thus mask the underlying deficit (92). However, patient awareness of the additional neural effort required to maintain performance is reflected in their low self-ratings of cognitive function. Hyperactivation has been consistently observed in association with aging and is greater in individuals with age-related neurodegeneration (97–99).

Resting state fMRI functional connectivity studies provide insight regarding the brain network’s organization in terms of parallel information processing. Some studies show diffuse hyperconnectivity in chemotherapy exposed breast cancer patients and survivors compared to controls (88, 100, 101) while others have shown hypoconnectivity (45, 102–106). Given the brain’s vast complexity, the interpretation of hyper- vs. hypo- activation/connectivity is difficult especially since both can exist simultaneously. Most observations have been made by comparing patients and controls. It is likely that there are subgroups of patients characterized by distinct patterns of brain abnormalities which may help clarify some of the inconsistent findings observed across studies.

Once a precision health CRCI diagnosis or prognosis is made, an intervention may be required. Currently, there are no standardized, evidence-based interventions specifically for CRCI. However, several pharmacological and behavioral/integrative interventions have been examined as potential treatments for CRCI. Pharmaceutical treatments have included psychostimulants (155, 156), nicotine patches (157), anemia medications (158), and anti-dementia medications (159). Nguyen and Ehrlich described multiple other drugs that could potentially be repurposed for treating CRCI (160). Preclinical studies have evaluated several novel agents including PAN-811, a ribonucleotide reductase inhibitor (161), mesenchymal stem cells (162–164), functionalized mitochondria (165), dual leucine zipper kinase (166), histone deacetylase 6 inhibitor (167), KU-32, for mitochondrial repair (168), and astaxanthin, an antioxidant (169), among others. However, these treatments have yet to be translated into clinical studies.

There have also been several studies examining behavioral (e.g., psychoeducational, cognitive behavioral therapy, compensatory strategies, cognitive training, cognitive rehabilitation), lifestyle (e.g., physical activity, diet, stress management), neuromodulation (neurofeedback, transcranial magnetic stimulation), and integrative (e.g., acupuncture, music, meditation, yoga) interventions for CRCI (153, 170–176). A meta-analysis of 29 randomized control trials reported that the best options among behavioral interventions for CRCI in descending order of efficacy were: meditation/mindfulness-based stress reduction, combined cognitive training with exercise, cognitive training, cognitive rehabilitation, exercise, cognitive behavioral therapy, qigong, supportive therapy, yoga, and acupuncture (170). Another meta-analysis of nonpharmacological interventions for CRCI reported significant effects on objective measures of attention, immediate recall, and processing speed and subjective cognitive function, depending on intervention type and mode of delivery (177).

Measuring outcomes in CRCI intervention trials based on symptoms alone may yield inaccurate results given the limited sensitivity and specificity these assessments have for CRCI and the lack of reliable impairment cutoffs. For example, several previous CRCI intervention trials have observed few if any significant changes in cognitive testing scores yet demonstrated significant effects on functional brain metrics in breast cancer survivors (178, 179). Stimulant medication trials for CRCI have shown very mixed results (180) despite anecdotal evidence that these medications can be helpful for certain patients with CRCI and fatigue. It is possible that a subgroup of patients benefits from these medications or that the outcome measures used in these trials were not sufficiently sensitive for detecting cognitive changes. Biotypes allow us to identify biologically based classifications that could be used as clinical endpoints for intervention trials and to determine more accurate cut off scores for classifying impairment from symptom assessments.

Animal models are not likely to be specifically useful for biotyping or prediction modeling since these methods rely on the patient’s individual biology and developing animal models for CRCI poses significant challenges due to the intricate nature of replicating the human disease (181). However, for optimal clinical utility, CRCI diagnostic and prediction models require further knowledge regarding which chemotherapies are associated with cognitive deficits. The biological effects exerted by each chemotherapy drug are extensive, and this complexity increases when combined with other medications. This cannot be ethically examined in human patients given that patients cannot be randomized to different treatment regimens. Preclinical models offer the advantage of assessing individual chemotherapeutic agents and elucidating their specific molecular mechanisms. By employing preclinical models, researchers can also better regulate variables such as age, sex, environmental factors, type of cancer, treatment types, and comorbidities [see review by Seigers et al. (182)]. Several recent studies have demonstrated cognitive and neurobiologic deficits in mice treated with doxorubicin, cyclophosphamide, or cisplatin chemotherapy (183–188). Potential mechanisms of CRCI examined in these studies include sphingosine-1-phosphate receptor 1 activation, brain derived neurotrophic factor levels, apolipoprotein E4 genotype, and neuroinflammation [see review by Gibson and Monje (189) for potential mechanisms]. However, few studies have compared different chemotherapy regimens or utilized combination chemotherapies and therefore, significant work is still required in this area.

Biotyping is a nascent field that warrants replication of results across several dimensions, which is known as deep validation (72, 190). Deep validation steps including applying an existing biotype algorithm to new, independent data without reclustering to determine if biotype expression remains the same as in the original data, clustering independent data with a naïve algorithm to determine if biotypes can be reproduced, and extending cross-sectional biotypes to longitudinal data (72). These methods aim to determine whether the existence of biotypes in a particular condition is reliable, if the original biotyping solution is generalizable, and if biotypes can predict different disease trajectories.

Most biotyping studies to date have employed fMRI neuroimaging. However, neuroimaging it is not currently standard practice for breast or other non-CNS cancers, it is costly and associated with contraindications such as electronic or magnetic biomedical implants, magnetic orthodontia, and claustrophobia, among others. Neuronal activity is associated with increased oxygen delivery and the difference, or contrast between diamagnetic oxygenated hemoglobin and paramagnetic deoxygenated hemoglobin is measured by MRI. This is known as the blood-oxygen-level-dependent (BOLD) contrast and is the standard method used in functional MRI (191). There are other non-invasive technologies that are used to measure brain activity including electroencephalogram and magnetoencephalography. However, functional near-infrared spectroscopy (fNIRS) is the most similar available technology to fMRI. FNIRS also measures the BOLD contrast but does so using near-infrared light. Whereas most biological tissues are transparent to near-infrared light between 700 and 900 nanometers, hemoglobin absorbs and scatters near-infrared light in this range. Oxygenated and deoxygenated hemoglobin absorb near-infrared light at different wavelengths, which constitutes the BOLD contrast for fNIRS measurement (192, 193). FNIRS is highly tolerant to motion and has no environmental restrictions, contraindications or known risks (194).

The primary limitations of fNIRS are that it can only measure cortical tissue and cannot penetrate to subcortical structures (194) and most commercially available fNIRS devices only measure frontal regions. However, it is important to point out that prefrontal cortex is involved in most brain functions (195–197). Compared to chemotherapy naïve patients, chemotherapy-treated patients with breast cancer consistently show significantly changed prefrontal structure and function (74, 79, 82, 85, 198–201). Altered prefrontal regions are among the most predictive of long-term cognitive status in patients with breast cancer (147, 148). Further, in patients with breast cancer, lower prefrontal-executive function is the single best predictor of medication non-adherence (202) and has been associated with decreased quality of life following chemotherapy treatment (203). Thus, prefrontal fNIRS has significant potential as a proxy for whole brain fMRI in the study of CRCI that could be more easily used in-clinic. Surprisingly, no studies to date have applied fNIRS to the study of CRCI.

The correspondence between liquid and neuroimaging biomarkers is another avenue of research for discovering potential proxy technologies that support biotyping. Several lines of evidence suggest that cancer and its therapies accelerate aging processes (77, 86, 204–206). Individuals with age-related neurodegeneration and related cognitive decline tend to demonstrate greater amyloid-beta peptide accumulation and tau hyperphosphorylation compared to cognitively normal individuals (207). Tau and amyloid-beta can be measured from serum or plasma using the commercially available, single-molecule array (SIMOA) assay (208–210). Several studies have demonstrated that blood-based measurements of amyloid-beta or tau using SIMOA are associated with age-related neurogenerative conditions (211–214) as well as other neurologic conditions (215, 216). We uniquely demonstrated that blood-based amyloid beta and tau are highly predictive of cognitive functioning in breast cancer survivors, in combination with age, cytokines, and anthropomorphic measures (217).

Inflammatory biomarkers have long been of interest to researchers studying CRCI. Objective measures of CRCI have been shown to correlate with some inflammatory cytokines (e.g., IL-1β, IL-6, TNF-α, CRP), other proteins (e.g., GM-CSF), neurodegenerative markers (e.g., Aβ − 42, Aβ − 40, tau), and methylation ratios (42). Subjective measures of CRCI have demonstrated both significant and nonsignificant correlations with cytokines (IL-4, IL-1β, IL-6, MCP-1, IGF-1) and no significant correlations with TNF-alpha or CRP (42), suggesting perhaps more complex and inconsistent relationships between inflammatory biomarkers and CRCI.

Circadian biomarkers that can be measured objectively with saliva and blood or biorhythms (e.g., actigraphy) and are implicated in numerous side effects seen in cancer including CRCI. Similarly, to inflammatory biomarkers, data on circadian markers and CRCI relationships have been inconsistent and additionally quite limited in cancer research. For example, a recent study by Ancoli-Israel and colleagues showed that disrupted circadian rhythms were associated with reductions in objectively measured neurocognitive function in their longitudinal study of women with breast cancer (218). However, other studies failed to find any association between circadian rhythmicity and cognition in women with metastatic breast cancer (219). We suggest more research is warranted on the relationships between liquid and rhythm biomarkers in CRCI.

Neurofilament light, NF-L, a marker of neuroaxonal injury, is another protein that has received recent attention in relation to neurodegenerative diseases, including Alzheimer’s dementia (220). Increased blood concentration of NF-L have also been detected in person with frontotemporal dementia (221), Huntington’s disease (222), and Parkinson’s disease (223), suggesting that NF-L may be a biomarker of neurodegenerative processes in general. Blood based assays using SIMOA to detect NF-L are commercially available but no studies of NF-L and CRCI have been conducted to date.

Another emerging biomarker of cognitive decline is known as neuron derived exosomes (NDE). Exosomes are extracellular vesicles that play an integral role in intracellular communication by conveying materials such as RNA and proteins between cells (224). Exosomes are released by a most cell types, including neurons, and can be detected in blood, saliva, and urine (225, 226). NDE’s appear to contribute to cognitive impairment by transporting toxins during pathological conditions (224). This is evidenced by findings of increased NDE levels in individuals with neurodegenerative syndromes (227, 228). Koh et al. provided a comprehensive overview of NDEs as potential biomarkers of CRCI (229).

CRCI research initiatives, including many funding opportunities, have emphasized the need for longitudinal studies as they determine patterns of individual cognitive change which can be more sensitive than a single static cognitive assessment. Longitudinal studies can also provide insight regarding the separate and combined effects of different anti-cancer treatments on cognitive function. However, large cross-sectional studies should not be neglected or discounted as they help establish reliable syndrome phenotypes (230) and are necessary for the initial steps in biotyping. Large multi-site, national studies or pooled studies are needed to validate neuroimaging biomarkers and biotypes of CRCI. Data sharing and pooling can be facilitated by online databases such as EBRAINS,¹ the Human Connectome Project,² and LONI,³ among others.

Finally, most CRCI studies to date have focused on breast cancer although growing evidence indicates cognitive impairment across cancer types (231). A critical component of future research will be the ability to demonstrate that biotypes can be applied to other cancer diagnoses to determine what CRCI is, agnostic to disease. Further research is required to examine biomarkers of CRCI across different cancer types.

CRCI is a prevalent condition that is not fully understood and, as a result, not diagnosed accurately, limiting treatment trials and implementation. Advanced statistical methodology, such as growth mixture modeling, allows for determination of latent subgroups using self-report and longitudinal data, offering significantly more precision in understanding cognitive trajectories and outcomes. However, the main limitation of symptom-driven subtypes classification remains lack of biological data. Neuroimaging can be used to identify brain biomarkers both individually and in clusters to define biological subgroups to diagnosis and predict outcomes. Like symptom-derived subtypes, biotyping accounts for variability in patient functioning by evaluating brain function to define the best taxonomy for patients but does so by utilizing the mechanistic information required for precision medicine.

SK: conception or design. SK, AH, SP, OP, and MW: acquisition, analysis, or interpretation of data, and drafting and revising the work. All authors contributed to the article and approved the submitted version.

This work was supported by funding from the National Institutes of Health (1R01CA226080, 2R01CA172145).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

The handling editor NS declared a past co-authorship with the author SK.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.