Letter to the Editor: Balancing efficacy and toxicity: The critical role of predictive models in colorectal cancer chemotherapy

Abstract

While the pioneering study by Liu et al published in World Journal of Gastroenterology masterfully addresses the critical challenge of predicting chemotherapy-induced myelosuppression, it also illuminates the necessary path for future research: The development of integrated models that concurrently predict a spectrum of common toxicities, such as gastrointestinal distress, neurotoxicity, and hand-foot syndrome, and crucially, link these predictions to efficacy outcomes. Personalized oncology aims to identify the regimen offering the maximal therapeutic benefit for each patient. A decision tool that flags both severe gut toxicity risk and a high probability of profound pathological response would provide clinicians with an invaluable, holistic framework. By expanding the present robust methodology into an integrated platform predicting multi-organ adverse events and tumor response, we can advance beyond isolated risk assessment and truly optimize the balance between treatment safety and anti-tumor activity, ensuring that therapeutic choices are as comprehensive and patient-centric as possible.

Key Words: Clinic-machine learning nomogram; Efficacy; Toxicity; Colorectal cancer; Chemotherapy

Core Tip: We propose expanding Liu et al’s interpretable machine learning-based nomogram beyond myelosuppression to an integrated efficacy-toxicity platform that simultaneously forecasts tumor response and common multi-organ adverse events. By embedding these dual predictions into a real-time, utility-based clinical decision support system, designed for practical clinical use, clinicians can instantly identify the regimen offering each patient the maximal therapeutic index, thereby optimizing personalized treatment decisions.

TO THE EDITOR

Liu et al’s study[1] published in World Journal of Gastroenterology for adeptly predicting the risk of chemotherapy-induced myelosuppression (CIM) in colorectal cancer patients is a scientific and pioneering work. By integrating multiple machine learning algorithms via a feature mapping approach, they have successfully bridged the critical gap between the high predictive accuracy of complex models and the clinical necessity for interpretability. With an impressive area under the receiver operating characteristic curve of 0.95 in the testing set, this model empowers clinicians to proactively manage one of the most common and debilitating dose-limiting toxicities in colorectal cancer treatment.

CIM ALONE IS NO LONGER ENOUGH

The successful prediction of CIM illuminates the next frontier in personalized chemotherapy: Developing integrated models that predict a spectrum of toxicities while crucially linking them to efficacy outcomes. CIM is but one facet of a patient’s tolerance to chemotherapy. Other frequent and equally impactful adverse events include severe gastrointestinal distress (diarrhea, nausea/vomiting), debilitating neurotoxicity, hand-foot syndrome, and hematological toxicities[2,3]. The clinical decision-making process weighs all these potential toxicities simultaneously (Table 1).

Table 1 Common toxicities of first-line colorectal cancer chemotherapy regimens that warrant integrated prediction.

Toxicity type	Occurrence of risk	Clinical impact	Associated regimens	Impact on decision-making
Chemotherapy-induced myelosuppression	High (FOLFOX); low (CAPOX)	Neutropenia, anemia, thrombocytopenia; risk of fever, infection, fatigue, bleeding	FOLFOX, FOLFIRI, CAPOX	Common reasons for dose delay, reduction, or use of growth factor support
Severe diarrhea	High (FOLFOX); medium (CAPOX)	Watery stools, dehydration, and electrolyte imbalance	FOLFOX, FOLFIRI, CAPOX	Frequent cause of dose modification or treatment interruption; requires proactive management
Peripheral neuropathy	Low (FOLFOX); high (CAPOX)	Numbness, tingling in the limbs, can progress to functional impairment (e.g., difficulty with fine motor tasks)	FOLFOX, CAPOX	Cumulative toxicity often leads to dose reduction or discontinuation of oxaliplatin
Hand-foot syndrome	Medium (FOLFOX); high (CAPOX)	Erythema, swelling, pain, blistering on palms and soles	FOLFOX, CAPOX	May necessitate dose reduction or interruption of capecitabine/fluorouracil

FOLFIRI: Fluorouracil, leucovorin, and irinotecan; FOLFOX: Fluorouracil, leucovorin, and oxaliplatin; CAPOX: Capecitabine and oxaliplatin.

CURRENT PREDICAMENT AND FUTURE RESEARCH DIRECTIONS

This field is already witnessing the emergence of predictive models for individual toxicities. For instance, pharmacogenetic testing for uridine diphosphate glucuronosyltransferase family 1 member A1 is routinely used to identify patients at high risk for irinotecan-induced severe neutropenia and diarrhea[4,5], while another research area is identifying genetic markers for oxaliplatin-induced neurotoxicity[6,7]. Similarly, models exist to predict the efficacy of chemotherapy. Studies have explored radiographic features, transcriptomic signatures, and clinical factors to forecast responses to neoadjuvant therapy or outcomes in metastatic settings[8-12]. Yet, these models often operate in isolation, creating a fragmented decision-making landscape. A clinician may be confronted with one model suggesting high efficacy for fluorouracil, leucovorin, and irinotecan (FOLFIRI) and the other predicting severe diarrhea from the same regimen. The CIM prediction model of Liu et al[1] indicates a moderate risk of myelosuppression. Synthesizing this multidimensional prediction into a single, coherent treatment plan that optimally balances efficacy against toxicity remains a profound clinical challenge. The value of a unified model lies precisely in its ability to illuminate this critical trade-off, rather than presenting it as a conflict.

Therefore, we propose that the ultimate goal is not merely to avoid a single toxicity, but to identify the treatment regimen with the most favorable therapeutic index for each individual patient. A unified model that could predict, for instance, both a patient’s high risk for severe diarrhea from irinotecan and their high probability of achieving a profound pathological response to an FOLFIRI regimen would provide an invaluable, holistic framework for clinical decision-making. This framework could be practically integrated into the clinical workflow, such as during pre-chemotherapy multidisciplinary team meetings or patient consultations, to visually compare the predicted efficacy-toxicity profiles of different regimen options. Such a model empowers clinicians by explicitly quantifying the competing dimensions of benefit and risk. In this scenario, a clinician might still justifiably choose FOLFIRI but intensify prophylactic anti-diarrheal measures and schedule more frequent follow-ups. Conversely, if the predicted efficacy is low and the toxicity risk is high, an alternative regimen would be strongly favored.

We envision the development of an integrated clinical decision-support platform that synthesizes these elements. Future research should build upon the robust methodology established by Liu et al[1]. The next step is to integrate a wider array of inputs to create a unified, multi-output predictive framework. This proposed platform would not be a simple aggregation of independent sub-models, but an integrated system where the interactions between features for different outcomes can be captured (Figure 1). Multi-task learning (MTL) architectures offer a powerful paradigm for developing such an integrated prediction platform. MTL models, particularly those based on neural networks, are designed to learn multiple related tasks (e.g., predicting various toxicities and efficacy) concurrently by sharing representations across tasks[13]. This approach can improve generalization for individual tasks by leveraging common underlying factors and inherently models the correlations between different clinical outcomes, moving beyond isolated predictions. MTL frameworks have been widely applied in oncology, providing a methodological precedent for our proposal[14-16].

Figure 1 Integrated efficacy-toxicity framework for precision first-line colorectal cancer chemotherapy. CA19-9: Carbohydrate antigen 19-9; CA125: Carbohydrate antigen 125; KRAS: Kirsten rat sarcoma viral oncogene homolog; BRAF: B-Raf proto-oncogene: Serine/threonine kinase; TP53: Tumor protein p53; UGT1A1: UDP glucuronosyltransferase family 1 member A1; GSTM1: Glutathione S-transferase mu 1; ABCC2: ATP-binding cassette subfamily C member 2; CYP3A5: Cytochrome P450 family 3 subfamily A member 5; GSTP1: Glutathione S-transferase pi 1; CRC: Colorectal cancer.

However, the development of this platform faces substantial challenges. Data requirements are paramount, necessitating large, curated, multi-modal datasets that harmonize clinical variables, genomic data, biomedical images, and longitudinal toxicity reports. The most pivotal challenge, however, lies in quantifying the “therapeutic index” for clinical decision support. This requires defining a clinically valid utility or net benefit function, where weights reflecting the relative clinical importance of different efficacy levels and toxicity severities are formally incorporated. Determining these weights is a non-trivial research endeavor in itself, likely requiring methods like discrete choice experiments involving oncologists, patients, and health economists to establish consensus[17,18]. Therefore, advancing this field demands not only robust predictive algorithms but also a parallel interdisciplinary effort to bridge predictive analytics with actionable clinical choice.

CONCLUSION

We again congratulate Liu et al[1] on their valuable contribution. Their work provides a solid methodological foundation and a powerful tool for managing CIM. By developing integrated models that concurrently predict toxicity and efficacy, we can move from isolated risk assessment to truly optimizing the balance between treatment safety and anti-tumor activity.