Navigating the landscape of personalized oncology: overcoming challenges and expanding horizons with computational modeling

Abstract

Objectives

We discuss challenges using computational modeling approaches for personalized prediction in clinical practice to predict treatment response for rare diseases treated by novel therapies using clinical oncology as an example context. Several challenges are discussed, including data scarcity, data sparsity, and difficulties in establishing interdisciplinary teams. Machine learning (ML), mechanistic modeling (MM), and hybrid modeling (HM) are discussed in the context of these challenges.

Materials and Methods

We present an HM approach, combining ML and MM techniques for improved personalized model estimation in the context of chimeric antigen receptor T-cell therapy for aggressive lymphoma.

Results

The HM approach improved the root mean squared error by $61.27\pm 23.21\%$ compared to using MM alone (MM: $2.36*{10}^5\mp 1.68*{10}^5$and HM: $9.57*{10}^4\mp 8.37*{10}^4$, where the units are in cells), computed from 13 patients included in this study.

Discussion

By exploiting the complementary strengths of ML and MM approaches, the developed HM method addresses common limitations such as data scarcity and sparsity in medical settings, especially common for rare diseases.

Conclusion

The HM techniques are likely required to overcome data scarcity and sparsity issues in broad medical settings. Developing these techniques requires dedicated interdisciplinary teams.

Introduction

The success of cancer treatment depends critically on our ability to predict which treatment choice will be successful. Incorrect therapy selection leads to treatment failure and unnecessary treatment-associated toxicities. However, in the case of novel treatments for a relatively rare disease, accurate treatment-response prediction is challenging because of data scarcity and sparsity.^1,2 Here, we define data scarcity as a limited number of observations or subjects available for analysis. Similarly, we define data sparsity as rare, missing, or incomplete data within single or repeated observations resulting from measurements collected at insufficient frequency. Data scarcity and sparsity problems often require new computational methods and approaches, raising an additional obstacle: the challenges of interdisciplinary collaborations.³ The required insights to develop methods for accurate and useful predictions are only possible when interdisciplinary team members understand each other’s perspectives, needs, and languages. A common informatics goal is to develop methods that are robust and generalize to other diseases with similar data limitations. For example, efficient use of computational methods for different cancer types could have broad implications, such as biomarker and risk factor identification, understanding the underlying dynamics of treatment-immune system-tumor microenvironment, and predicting the treatment option with the highest likelihood of success.^4–7 In this Perspective, we discuss recent advances that can potentially support improving patient outcomes through new methods for accurate prediction of treatment responses for novel treatments of relatively rare cancers. We also discuss fundamental roadblocks that are limiting the use of these approaches in clinical oncology. Finally, we offer some potential solutions. We illustrate these concepts with an example of computational modeling in cancer immunotherapy.

Background on computational predictive modeling

Because of data limitations and system complexity, progress in personalized treatment outcome prediction for oncology patients may require multiple computational modeling approaches that combine models that encode knowledge of biological systems via mechanistic modeling (MM) and model-free machine learning (ML). Figure 1 gives a conceptual overview of such a predictive modeling process. Predictive modeling requires two high-level choices: selecting a class of models and an inference method. Two broad classes of models include (i) mechanistic models^8,9 that use existing knowledge of the biological system estimated using data assimilation (DA),^10,11 a framework that is often known for creating a digital twin¹² because the model is constructed to mimic the mechanics of the observed system and is continuously compared and updated with data as they arrive, and (ii) parameterized families (eg, linear, Gaussian) or families of non-parametric functions (eg, neural networks, transformers) estimated with ML methods.¹³ Inference methods are wide-ranging, including Kalman filtering^14,15 optimization, variational techniques,¹⁶ stochastic methods such as Markov chain Monte Carlo,¹⁷ and are sometimes tied to the model class being estimated.

Figure 1.

Summary of mathematical modeling techniques. Depending on the task, inputs could cover a wide range of variables: patient-specific, disease-specific, and drug-specific variables. Then, based on the data-related constraints and the purposes, the most appropriate modeling approach is selected. This choice also directs the choice of a suitable inference scheme. As a result of the computational framework created by all the components mentioned so far, one or a few of the wide range of outputs could be achieved.

Predictive methodology choice is determined by the context and constraints of the clinical problem. MM and ML methods have distinct strengths and limitations. MM can perform well with small or sparse datasets and is typically estimated separately for each individual.^18,19 It leverages known relationships between variables, can infer unmeasured quantities, but requires an existing model or sufficient system knowledge to develop one. ML, by contrast, generally requires large datasets and operates at the population level, applying generalized patterns to individual predictions. While ML can capture complex patterns in high-dimensional data, it lacks built-in knowledge of the underlying system and may struggle in data-sparse settings. To overcome the limitations of each, recent work has focused on developing hybrid MM–ML frameworks that integrate their complementary strengths.^20–28

Key challenges of using computational predictive modeling in oncology

In the context of clinical cancer research, both ML and MM approaches have been used for a variety of applications. ML models have been used in oncology for risk assessment,^29–34 lesion detection and characterization,^35–37 therapy response prediction,^38–43 in the design of new drugs, biologics, and cellular immunotherapies.⁴⁴ MM have been used to describe the time-dependent dynamics of CAR T-cells and tumor cells,^45–48 revealing pharmacokinetic/dynamic impacts on adverse events or even the dynamics of cytokines.⁴⁹ These cases are excellent examples of conceptualizing the challenges discussed in the introduction that have limited the development and implementation of personalized prediction models in clinical oncology, including data sparsity, scarcity, and lack of interdisciplinary collaboration.

Data sparsity and its limitations in predicting cancer treatment outcomes

In the clinic, histologic and radiographic assessments are used, and panel-based sequencing data are increasingly common. However, these clinical data are often incomplete or missing due to technical limitations of testing (eg, limited biopsy specimen, poor quality sample for sequencing), lack of patient follow-up, limited health care system resources, lack of insurance coverage, and the proprietary nature of some drug-specific data. However, a larger problem is the availability of biologically meaningful information to clinicians. Cancers are biologically complex, heterogeneous, and dynamic diseases that evolve genetically over time and interact with the patient’s endogenous tissues, immune microenvironment, microbiome, exogenous environment, and prior treatments.⁵⁰ Given this biological complexity, routinely collected data for cancer patients capture only a small fraction of the biologically relevant information that could potentially be used for prediction and personalized treatment planning. Although panel-based genomic sequencing of tumors has become common, the significance of the results is often not known or actionable.^51,52 It is likely that incorporating more comprehensive and granular data such as whole genome sequencing, transcriptional profiling, tumor microenvironment characteristics, and immunophenotyping of the tumor and endogenous immune system would allow improved prediction.^53–58 Finally, because clinical assessments are commonly performed at relatively infrequent intervals (eg, cross-sectional imaging every 3-6 months), the number of observations per patient is limited.

Addressing data scarcity in computational predictive modeling of cancer treatment outcomes

Publicly available data repositories (eg, UK Biobank, All of Us Biobank) are potential sources of granular data for thousands of subjects but often lack key context-specific variables. The sensitive nature of personal medical data has posed a significant challenge to accumulating the large data sets frequently needed for ML analysis. In addition, the relative rarity of many cancer subtypes limits data set sizes. Sophisticated deep learning methods, with millions of parameters, require massive data sets. Even shallow nonlinear ML methods require large data sets unless they are severely regularized.^13,59

Diverse approaches are being considered to address these data limitations. First, newer statistical methods are being developed to address data missingness. These include advanced statistical and deep learning-based imputation methods.^60,61 MM is well-suited to overcome data scarcity and sparsity issues by exploiting a mathematical representation of the underlying biological system rather than learning by identifying patterns in data alone. While mechanistic models are not completely insensitive to sparse data, they can often accommodate missing data more efficiently than ML models. Combining ML and MM approaches toward the development of predictive models is a novel way to potentially overcome data sparsity limitations. Being relatively new, such hybrid modeling (HM) approaches have started to be applied in different medical settings and are shown to improve prediction accuracy in several cases, such as prediction of metastatic relapse of breast cancer,²³ intensive care unit (ICU) mortality prediction,⁶² risk assessment for stroke,⁶³ and forecasting blood glucose values of type 1 diabetes patients,⁶⁴ type 2 diabetes and ICU patients.^65,66

Another strategy to overcome data scarcity is federated learning.^67–69 This is a decentralized, collaborative ML technique in which algorithms are trained on local data sets, with the resulting models subsequently aggregated to generate an updated central model.^70,71 There are a few examples of federated learning in oncology,^71–73 but application challenges remain. Federated learning provides the ability to use different local data sources by facilitating data privacy, a common concern in medical settings. Although the models are trained locally, a centralized model is updated based on information from local model training, incorporating a large amount of additional data from a wide range of local sources, boosting accuracy without directly accessing the local sensitive data. Finally, although synthetic data sets have been proposed as one solution to data scarcity in biomedical research, the fidelity of such data and a regulatory framework for their use is currently lacking. All these techniques address the data sparsity challenges from different perspectives. Combining these methods could support vast potential improvements in prediction accuracy. For example, using synthetic data generated from real-world data can support the development of accurate hybrid models. The prediction accuracy can be further improved if these hybrid models are employed within federated learning algorithms, being updated once new local data become available, and improving the central model.

Overcoming challenges in cross-disciplinary collaboration for personalized oncology

Interdisciplinary collaboration is challenging and is a critical barrier to advancing personalized oncology. Fundamental differences exist between how a clinical decision-making problem (eg, the likelihood of success of a therapy) is approached by a clinician versus a computational method. A clinician’s approach accounts for many factors that may not be quantifiable (eg, the likelihood of patient compliance) or present as measured data and, therefore, is challenging to incorporate into a modeling framework. In contrast, computational tools can be helpful in extracting intricate patterns from complex data, patterns likely to be overlooked or incomprehensible to clinicians.⁷⁴

Development of clinically useful computational models requires understanding (i) the day-to-day realities of patient care and the practical considerations of implementing novel clinical tools, (ii) the technical and computational knowledge to develop robust, accurate, and validated tools, and (iii) translation via the design of decision-aid tools that can inject the computational model-based predictions into the clinical decision-making process. This requires the unification of deep clinical, health care process, computational, and human-computer interaction insights. The clinical world is beginning to recognize the critical importance of such collaborative team efforts in advancing these methods in routine clinical practice.^75,76

Methods

An illustrative example from large B-cell lymphoma

To illustrate the limitations induced by data scarcity, sparsity, and of using only MM or ML and the potential utility of personalized HM in oncology, we consider the case of relapsed and refractory large B-cell lymphoma (LBCL).⁷⁷ The development of chimeric antigen receptor (CAR) T-cell therapy has led to a paradigm shift in the treatment of patients with relapsed LBCL. However, despite the significant improvement in outcomes seen with this therapy compared to historical treatment options,^78,79 only 40% of patients derive long-term benefits. Early and accurate prediction of treatment outcomes after CAR T-cell therapy infusion would help clinicians strategize the next steps for better health outcomes, especially if the treatment is predicted to fail. Therefore, we aimed to estimate the personalized CAR T-cell and cancer cell dynamics for LBCL patients within fifteen days after the CAR T-cell infusion based on routine clinical data, a challenging problem with data sparsity and scarcity. An accurate estimation of these dynamics could elevate our physiological understanding of the underlying system and could be used for treatment outcome prediction. Here, we merely focus on investigating the usability of MM with extremely sparse routine clinical data and the illustration of one possible approach for HM.

Typical large B-cell lymphoma data

Our data set consisted of routine clinical data from 13 patients, whose clinical data are shown in Table 1. LBCL data with CAR T-cell data are relatively rare, and while our data are representative of such data sets, the limited population size represents a common severe limitation. This study was reviewed and approved by the Colorado Multiple Institutional Review Board (protocol number: 21-4232). We used total metabolic tumor volume (TMTV) to estimate the number of lymphoma cells present before and after CAR T-cell infusion.⁸⁰ The TMTV measurements are sparse, totaling one to three measurements for each patient, only one of which can be used to create the personalized predictive model. Furthermore, the pre-infusion TMTV measurement, which served as an approximation to tumor volume at the time of CAR T-cell infusion, ranged from 0 to 1942.5 cm³ across patients and was collected up to three months prior to CAR T-cell infusion. One patient had no detectable TMTV on the PET scan. For modeling purposes, we assigned this patient a TMTV of 0.1 cm³, the smallest volume reliably detectable by a PET imaging. The number of infused CAR T-cells was known, and the expansion of CAR T-cells was characterized using the absolute lymphocyte count (ALC) measured over the first fifteen days following CAR T-cell infusion.⁸¹ The mean and standard deviation of the number of ALC measurements were $8.8\mp 1.8$, with a range $\left[\mathrm{5,11}\right]$ across all patients, highlighting the sparsity of the data. Moreover, although the sample size will grow, the complexity and exploratory nature of using CAR T-cell therapy is likely to remain small until it becomes a more common cancer therapy, as is the case for all new promising treatments.

Table 1.

Summary statistics for all 13 patients, with individual results shown for patients 3, 4, 7, and 13, who were randomly selected and are representative of the cohort’s range of outcomes.^a,b

Variable	Patient 3	Patient 4	Patient 7	Patient 13	Summary for all 13 patients
Age at CAR T-cell infusion	75	75	65	53	mean: 62.2, stdev: 11.8, median: 65, iqr: 16.3, min: 36, max: 75
Sex	F	F	M	M	Female: 4, Male: 9
Race	White	White	White	American Indian	American Indian: 1, Asian: 1, White: 11
Lymphoma subtype	LBCL	LBCL	LBCL	LBCL	LBCL: 8, HGBL: 5
Ki67(%)	55	85	85	55	mean: 77.7, stdev: 17.8, median: 85, iqr: 27.5, min: 40, max: 95
Pre-CAR T infusion TMTV (${\mathrm{cm}}^3$)c	196.4	1942.5	450.9	1.6	mean: 273.9, stdev: 527.3, median: 74.6, iqr: 323.0, min: 0, max: 1942.5
Number of ALC measurements used in modelingd	10	12	10	11	mean: 8.8, stdev: 1.8, median: 9, iqr: 2, min: 5, max: 11
Best response to CAR T-cell therapy	CR	PD	PD	CR	Complete remission: 7, Progressive disease: 6
Time from CAR T-cell infusion to the last follow-up (months)	37.0	2.3	31.1	12.7	mean: 19.8, stdev: 12.7, median: 18.0, iqr: 23.2, min: 2.3, max: 39.3
Progression at last follow-up	No	Yes	Yes	No	Yes: 7, No: 6
Deceased at last follow up	No	Yes	Yes	No	Yes: 7, No: 6

^aThese four patients are included to provide concrete examples of variability in response and to illustrate the specific gains achieved through personalization, complementing the overall summary statistics.

^bALC: absolute lymphocyte count; CAR: chimeric antigen receptor; CR: complete remission; DLBCL: diffuse large B-cell lymphoma; HGBL: high-grade B-cell lymphoma; PD: progressive disease; TMTV: total metabolic tumor volume.

^cUsed in the model-based predictive algorithm.

^dThe number of ALC measurements is not directly used in the modeling, but the ALC measurements were used. The rest of the variables are included to provide context and baseline information about the patient cohort of this study.

Hybrid modeling approach

We used the following mechanistic model⁸² representing the dynamics and interaction of CAR T-cells and cancer cells.

$$\frac{dX}{dt}=\rho X\left(1-\frac{X}{K}\right)-\mathrm{ }{\kappa }_{1}XY,$$

$$\frac{dY}{dt}=\mathrm{ }{\kappa }_{2}XY-\theta Y,$$

where $X\left(t\right)$ and $Y\left(t\right)$ are the cancer cell and CAR T-cell counts at day $t$, respectively, $\rho$ is the cancer cell net growth rate (1/day), $K$ is the carrying capacity (cell), ${\kappa }_1$ is the CAR T-cell killing rate (1/(day*cell)), ${\kappa }_2$ is the net rate of proliferation and exhaustion of CAR T-cells when stimulated by cancer cells (1/(day*cell)) and $\theta$ is the CAR T-cell death rate (1/day).

Model estimation requires estimating unknown model parameters using individual patient data. In this setting, data include only one TMTV measurement and ALC measurements to be used for model fitting. Even for MM, model estimation is difficult when data are this sparse because the response to therapy is entirely missing in these data. Additionally, CAR T-cell data are noisy as we use surrogate measurements (ALC). Instead of exact predictive accuracy, we are focusing on solving the problem of limiting the model parameter space such that accurate solutions can potentially be found. This is an important problem because solving for optimal parameters in nonlinear optimization requires defining a feasible region or the set of parameters to search over. If the feasible region is too large, estimating optimal parameters often fails, a particularly acute problem with sparse data. To address this problem and achieve accurate, personalized prediction, we must guide optimization algorithms using ML toward effective solutions.

As a baseline, we estimated the model for each patient by using only their data over 15 days after CAR T-cell infusion. This window was selected because ALC measurements are considered reliable surrogates for CAR T-cell expansion only during this early phase, due to the transient effects of lymphodepletion.^81,83 Beyond this period, the relationship between ALC and CAR T-cell levels becomes less predictable. Although our long-term goal is to forecast tumor burden and treatment outcomes over extended time horizons, this study focuses on the early phase to demonstrate the feasibility of personalized, hybrid modeling approaches. We followed the standard optimization approach: estimating parameters by minimizing the least squares error with constrained optimization, where we defined the feasible region in the standard way (no ML). Without limiting the feasible region, sparse data and the complexity of the underlying system resulted in poor parameter estimation reflected by the model output (blue curve) and data (red circles) in the left-hand panels of Figure 2. We probed the impact of guiding the optimization algorithm using a combination of clinical and computational insight to mitigate this issue by searching parameters over a narrower feasible region, which must be identified for each patient before personalized model estimation. In these plots, the blue crosses represent pointwise simulations, model outputs evaluated only at the time points where data were available, while the blue curve represents the continuous simulation over the full time window. Pointwise and continuous values are identical at data time points; we include both for clarity in comparing observed data and model fit. We then generalized the method for computing a personalized feasible region identification using ML, hybridizing MM with ML. We took the following steps for this HM approach:

Figure 2.

We show data and model output for four patients estimated using the original (panels A, C, E, and G) and personalized (panels B, D, F, and H) feasible regions for two patients. In each plot, the red circles show data, the blue crosses show model output at the time of data points, and the blue curve shows the continuous model output. The term continuous simulation refers to evaluating the model across a fine time grid, independent of when the actual data were collected, whereas pointwise simulation refers to model outputs evaluated only at the time points corresponding to the observed data. As such, pointwise simulation values are identical to the continuous simulation values at those specific time points. We included the pointwise simulation to facilitate visual comparison between the model output and the observed data. Model parameters were estimated by minimizing the least squares distance between the pointwise simulation and the observed data. The visual overlap between pointwise and continuous simulation at the data points is complete by design; their difference lies primarily in presentation and purpose: the continuous curve helps visualize the model’s dynamic behavior over time, while the pointwise markers aid in directly comparing model predictions to available measurements. The plots show that using personalized feasible regions significantly improved model fit for all patients.

1. Based on clinical and computational insights, we constructed the search space for each parameter by combining literature-reported values, simulation results, and expert clinical input. For instance, the tumor growth rate (1/day) varies widely across cancer types depending on aggressiveness, with literature suggesting a range from 0.001 to 0.5 as broadly representative.⁸⁰ We found the range [0.01,0.1] to be appropriate for our cohort of 13 patients, balancing biological realism with sufficient flexibility for patient-specific variation. This approach was applied to other parameters to ensure that the model remained physiologically and clinically grounded. Notably, simulations revealed substantial variability in the parameter ${\kappa }_1$ (CAR T-cell killing rate) across patients. To obtain reasonable model fits, we found that the model required patient-specific feasible range for ${\kappa }_1$, while keeping the feasible ranges for all other parameters fixed. However, the selection of the feasible region that results in the best model fit requires automation to be practical in real-world clinical settings.

2. We estimated the model using each of these feasible regions separately for each patient.

3. For each patient, we automated the model selection by choosing the model resulting in the smallest mean squared error (MSE) between CAR T-cell count data and model output. We call the feasible region resulting in the optimal model the personalized feasible region. The automated selection of the optimal model at this step is the ML piece of this hybrid methodology.

What is gained by taking an interdisciplinary, MM-ML hybridized approach

To demonstrate the gains of the HM approach, we will present in-depth results from four randomly selected patients (Patients 3, 4, 7, and 13), who are also representative of the overall group of 13. To pull apart the HM gains, we use the root mean squared error (RMSE) computed using CAR T-cell count data (obtained from ALC measurements) and model outputs estimated using the original and personalized feasible regions, and the percentage improvement in RMSE when the HM approach is used compared to using the MM alone. These computational results for all 13 are shown in Table 2. In Table 2, we observe that the personalized feasible regions obtained automatically by hybridizing MM with ML reduce the estimation error for all patients except one. Notably, we did not conduct any ML (alone) estimation with these data because the data sets were too small and led to model estimates that were severely overfit, unreliable, or inestimable, depending on the method chosen. The HM approach demonstrates more accurate performance compared to MM and ML used alone. As an aside, patient 8 has a larger error compared to all the other patients. A likely source of this error is that this patient was delivered a smaller (one order of magnitude smaller) dose of CAR T-cells than all the other patients. Surely, there are other sources of estimation error, and more work is required to develop methods for discovering and understanding the impact of this difference on the model estimation error. Figure 2 shows the model outputs obtained by the original and personalized feasible regions for the four patients selected to show the model’s ability to represent varying system dynamics. We observe similar changes in the estimation results for all patients. The plots on the left-hand panels of Figure 2 show that using the original feasible region with sparse data results in estimates not reflecting the true dynamics. However, the true dynamics can be obtained by constraining the optimization algorithm in a patient-specific way, as shown on the right-hand panels. Figure 2 also shows that the model can represent varying system dynamics depending on the estimated parameters, eg, slower response to therapy (panel B), a rapid response resulting in eradication of cancer cells (panel D), initial response followed by resistance (panel F), and delayed response to therapy (panel H). This example demonstrates the substantial prediction accuracy gains that could be achieved when hybridizing MM with ML in the context of sparse data. Using more sophisticated ML techniques with the investigation of additional feasible regions once larger data sets are available could enhance the model fits shown in Figure 2. It is important to note that here, our goal was to present how HM can improve model fits and potentially prediction accuracy. Complete internal and external model validation requires a less limited data set with a larger population, which is a limitation of the current study that will be addressed in future studies.

Table 2.

We show the root mean squared error (RMSE) for each patient with model output computed by the original and personalized feasible regions and the percentage improvement in the RMSE when the hybrid modeling is used.^a

Patient	RMSE with original feasible region	RMSE with personalized feasible region	% improvement in RMSE
1	${1.70*10}^5$	${0.60*10}^5$	64.80
2	${2.35*10}^5$	${0.72*10}^5$	69.42
3	${1.23*10}^5$	${0.32*10}^5$	74.14
4	${2.39*10}^5$	${0.52*10}^5$	78.35
5	${0.61*10}^5$	${0.39*10}^5$	36.22
6	${2.06*10}^5$	${0.92*10}^5$	55.08
7	${2.51*10}^5$	${0.24*10}^5$	90.42
8	${3.21*10}^5$	${2.35*10}^5$	26.69
9	${0.80*10}^5$	${0.26*10}^5$	67.94
10	${3.05*10}^5$	${0.89*10}^5$	70.69
11	${0.87*10}^5$	${0.17*10}^5$	80.81
12	${4.00*10}^5$	${2.39*10}^5$	40.12
13	${2.40*10}^5$	${1.38*10}^5$	42.35

^aThe unit for the RMSE is CAR T-cell count.

Discussion

Limitations

A key limitation of this work is the small cohort size (13 patients), which constrains the generalizability of the specific findings and parameter estimates. The feasible regions identified were tailored to this limited group and may not directly apply to larger or more diverse patient populations. However, the methodological pipeline was designed with scalability in mind. The automated selection process is capable of adaptively identifying patient-specific feasible regions and optimal parameter fits, which supports broader application to larger cohorts in future studies. Expanding the approach to more diverse datasets will be important for validating its robustness and clinical utility.

Another limitation is that it does not account for the variability in timing between the last available TMTV measurement and the CAR T-cell infusion. We used the most recent pre-infusion TMTV value as a surrogate for tumor burden at the time of therapy, regardless of the actual time gap. This simplification may affect the accuracy of model fits, particularly for patients with rapidly proliferating disease, where tumor burden can change meaningfully even within short time frames. Incorporating this timing difference into the modeling, such as by forward-simulating tumor growth using proliferation markers like Ki67, could enhance the accuracy of tumor burden estimation and improve the clinical relevance of the approach in decision support settings. We view this as an important direction for future refinement.

Challenges of integrating hybrid modeling into clinical decision support systems

HM has the potential to enhance clinical decision-making by integrating various data sources and leveraging the complementary strengths of ML and MM methods. However, the integration of these models into clinical workflows requires addressing several key considerations. First, prior to incorporation into electronic health record (EHR) systems for clinical decision support (CDS), these models must undergo rigorous validation to ensure reliability and accuracy. Equally important is the way model outputs are presented to clinicians, ensuring the information is clear, actionable, and relevant to support informed decision-making. This requires careful consideration throughout the entire process—from model development to the implementation of a CDS tool—through collaboration with experts in human-computer interaction (HCI). Furthermore, the end users of these CDS tools need to be trained to interpret model-based recommendations and provide feedback for continuous system refinement, similar to what is done when new laboratory measurements become available. The successful deployment of reliable and actionable model-based CDS tools will require collaboration among interdisciplinary teams, including experts in healthcare, data science, HCI, and regulatory compliance, to address these various challenges effectively.

Data privacy and computational feasibility in clinical practice implementation

The hybrid modeling example we present here is computationally efficient, using minimal resources and individual patient data for personalized predictions. The latter feature ensures privacy by keeping protected health information (PHI) within the secure EHR system. This approach’s low computational demands allow for easy integration into existing workflows without significant changes or additional resources. It addresses two critical challenges in leveraging EHR data for computational modeling: it enables seamless incorporation into the EHR system and maintains data privacy by preventing PHI from leaving the secure environment, ensuring compliance with privacy regulations. Overall, it offers a practical solution for integrating advanced models into healthcare without compromising data security.

Enhancing data integration, scalability, and Real-World implementation with hybrid modeling

Hybrid modeling approaches can enhance data integration, scalability, and real-world implementation. For example, the methodological pipeline we develop here facilitates the incorporation of diverse data sources—clinical, genomic, and treatment-related—enabling accurate, personalized predictions. Moreover, its computational efficiency allows scalability to larger populations and diverse medical contexts, making it suitable for broader clinical applications without increasing complexity. Additionally, this example approach preserves data privacy by keeping sensitive information within secure EHR systems, and its simplicity ensures seamless integration into existing clinical workflows, addressing common implementation challenges.

Positioning within the broader informatics landscape

The methodological pipeline we present here is an example of how hybridizing MM and ML extends the framework of predictive modeling in oncology to handle data sparsity and improve prediction accuracy. This approach complements existing informatics methods by addressing key challenges in personalized treatment, including data privacy and computational tractability. Additionally, our hybrid modeling approach aligns with trends in other domains where similar issues of sparse data and the need for personalized predictions exist. For example, hybrid modeling has shown promise in healthcare domains such as ICU mortality prediction⁶² and risk assessment in stroke patients.⁶³ By situating our methodology within these established frameworks, we demonstrate its potential to enhance predictive modeling efforts not only in oncology but also in other healthcare settings where data limitations are a significant challenge. This positions our approach as both a novel and complementary advancement in the field of medical informatics.

Generalizability to other medical settings

Although the hybrid modeling example we present here was developed for personalized prediction of CAR T-cell therapy outcomes in LBCL patients, the methodological pipeline is generalizable to other medical contexts. Broadly, the goal is to use clinical insight and ML to reduce the model estimation space—the feasible region—such that model estimation is possible with minimal data, a common challenge across many medical settings.^28,84,85 Our pathway to address this goal is an interdisciplinary pipeline with three steps. The first step surfaces, discovers, and uses clinical insight and thought processes to identify the potential feasible regions, effectively creating the outlines of an algorithm. The second step is to mathematize this clinical reasoning process into an ML form that can be automated to select the most appropriate, narrowest feasible region for each patient. The third step is the development of the hybridized ML-MM algorithm to forecast treatment outcomes. While different diseases require different physiological models, this framework provides a systematic method to identify personalized feasible regions for optimal parameter estimation and model fitting across diverse clinical settings.

Conclusion

The development and use of computational modeling for clinical decision support in precision oncology demands overcoming both computational and social challenges, particularly in building and using interdisciplinary teams. In our example, data scarcity and sparsity rendered the use of ML and MM approaches alone unstable and inaccurate. An interdisciplinary team was necessary to hybridize MM with ML, guided by clinical insights, to develop an algorithm to achieve accurate and personalized model fits. Understanding the specific needs of clinical and computational fields required extensive discussions and a thorough literature review by all team members. Our conclusion is that building effective interdisciplinary teams requires dedication from team members to understand field-specific needs for the development of useful clinical decision support tools that can improve cancer treatment outcomes. Moreover, a deep understanding of the clinical context is crucial to selecting or developing the right computational approaches, addressing field-specific limitations, and enabling the use of computational modes in precision oncology and other medical fields.

Author contributions

Melike Sirlanci (Conceptualization, Methodology, Software, Visualization, Writing—original draft, Writing—review & editing), David Albers (Conceptualization, Methodology, Supervision, Writing—review & editing), Jennifer Kwak (Data curation), Clayton Smith (Supervision, Writing—review & editing), Tellen D Bennett (Supervision, Writing—review & editing), and Steven M. Bair (Conceptualization, Methodology, Visualization, Writing—original draft, Writing—review & editing)

Funding

This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Conflicts of interest

None declared.

Data availability

The data underlying this article were provided by the University of Colorado Health Cancer Center under license/by permission. Data will be shared on request to the corresponding author with the permission of the University of Colorado Health Cancer Center.