Purpose/Objective(s) Acute hospitalization during or after cancer treatment negatively impacts quality of care and causes significant patient morbidity. In patients receiving radiation therapy (RT) for any malignancy, we hypothesized that a machine learning approach would enable prediction of hospitalizations during or shortly after RT. Materials/Methods We analyzed 35,810 courses of RT provided to treat a cancer diagnosis at a large multisite academic department in a major metropolitan environment (all cancer patients treated with RT between 8/1999—1/2022 regardless of disease site or treatment intent). Approximately 150 clinical/treatment variables were extracted and processed into analytic format by a proprietary oncology analytics platform connected to the Electronic Medical Record System (inpatient and outpatient), Oncology Information System, and PACS. Such variables included past medical history, recent laboratory data, prior systemic cancer therapies, prior RT, recent hospitalization, and RT details. A patient was labeled as having suffered an acute hospitalization if they had an encounter classified as inpatient or emergency during RT or in the 30 days following completion of RT. A machine learning model was trained on a subset (75%) of the data reserved for training (26857 cases). Five-fold cross-validation was used to select model type and hyperparameters, using area under the ROC curve (AUC) to measure performance. The final model was evaluated for accuracy, AUC, precision, recall, and F1 score on an independent test set (25%) excluded from the training process (8953 cases). Model calibration was assessed by visual inspection of calibration plots. An AUC>0.70 was considered clinically valid. Results Among the 35,810 courses of RT, the incidence of acute hospitalization was 9.1% (9.2% training set; 9.0% test set). Model performance on the test set is shown in Table 1. Variables deemed to be significant predictors for hospitalization included recent lab values (PT INR, sodium, potassium) recent hospitalization (within 30 days prior to RT start), and patient age Conclusion In cancer patients undergoing RT, a machine learning model identified patients at risk of 30-day hospitalization. Predictive analytics may be a key tool to help providers identify high-risk patients and optimize interventions, while improving quality and value of care.

PURPOSE:Patients with gastrointestinal (GI) cancer frequently experience unplanned hospitalizations, but predictive tools to identify high-risk patients are lacking. We developed a machine learning model to identify high-risk patients.METHODS AND MATERIALS:In the study, 1341 consecutive patients undergoing GI (abdominal or pelvic) radiation treatment (RT) from March 2016 to July 2018 (derivation) and July 2018 to January 2019 (validation) were assessed for unplanned hospitalizations within 30 days of finishing RT. In the derivation cohort of 663 abdominal and 427 pelvic RT patients, a machine learning approach derived random forest, gradient boosted decision tree, and logistic regression models to predict 30-day unplanned hospitalizations. Model performance was assessed using area under the receiver operating characteristic curve (AUC) and prospectively validated in 161 abdominal and 90 pelvic RT patients using Mann-Whitney rank-sum test. Highest quintile of risk for hospitalization was defined as "high-risk" and the remainder "low-risk." Hospitalizations for high- versus low-risk patients were compared using Pearson's χ2 test and survival using Kaplan-Meier log-rank test.RESULTS:Overall, 13% and 11% of patients receiving abdominal and pelvic RT experienced 30-day unplanned hospitalization. In the derivation phase, gradient boosted decision tree cross-validation yielded AUC = 0.823 (abdominal patients) and random forest yielded AUC = 0.776 (pelvic patients). In the validation phase, these models yielded AUC = 0.749 and 0.764, respectively (P < .001 and P = .002). Validation models discriminated high- versus low-risk patients: in abdominal RT patients, frequency of hospitalization was 39% versus 9% in high- versus low-risk groups (P < .001) and 6-month survival was 67% versus 92% (P = .001). In pelvic RT patients, frequency of hospitalization was 33% versus 8% (P = .002) and survival was 86% versus 92% (P = .15) in high- versus low-risk patients.CONCLUSIONS:In patients with GI cancer undergoing RT as part of multimodality treatment, machine learning models for 30-day unplanned hospitalization discriminated high- versus low-risk patients. Future applications will test utility of models to prompt interventions to decrease hospitalizations and adverse outcomes.

To train a clinically valid predictive model for overall survival at 6 months, 1 year, and 3 years following radiation of brain metastases. A retrospective analysis of 915 patients treated with radiation therapy (RT) for brain metastases at one institution between 2011-2018 was performed. For each patient and RT course, 22 clinical and treatment variables were automatically extracted using an oncology analytics software platform integrated with our Electronic Medical Record System, Oncology Information System, Treatment Planning System, and Cancer Registry. All data was then de-identified in accordance with HIPAA safe harbor methods. Variables included patient demographics, tumor characteristics, prior treatment history, performance status, and RT treatment details. Overall survival (OS) status at six months, one year, and three years after initiation of treatment was extracted using the cancer registry as the primary system of record. Random forest, gradient boosted decision trees, and regularized logistic regression models were trained on a subset of RT courses consisting of 80% of the overall dataset. Five-fold cross-validation was used to select model type and tune model hyperparameters, using area under the ROC curve (AUC) as the scoring function. The best performing model (as determined by AUC on the training set) for each outcome was then evaluated on an independent test set consisting of 20% of the overall dataset. An AUC threshold of .70 was set as our definition of clinically valid. Of the 915 patients, 57.3% were female, the median age was 64 years (IQR: 56-71 years), the median KPS score was 70 (IQR: 60-80), the median volume of the PTV was 8.476 cc (IQR: 3.2135-19.75), and the median RT dose to the PTV was 21 Gy (IQR: 20-30 Gy). The top recorded primary cancer disease systems were thoracic (54.3%) and breast (17.1%). Incidence of each outcome in the training set and in the test set is show in Table 1. Performance of the best predictive model for each outcome is shown in Table 1. The best model type for all three outcomes was regularized logistic regression. In this study, we demonstrated that clinical valid predictive models of OS can be trained on data extracted automatically from a variety of institutional sources. The best performing model in each case leveraged regularized logistic regression and model performance degraded with more time elapsed from treatment. Quantitative prognosis tools such as these, in the hands of qualified physicians, have the potential to guide treatment decisions and inform end-of-life discussions. Additional efforts to validate these models prospectively on a multi-institutional dataset are necessary prior to introduction into clinical practice.Abstract 3786; TableIncidence of event and model performance for OS at each time interval.OutcomeTraining set incidenceTraining set AUCTesting set incidenceTesting set AUC6-month OS58.5%0.712459.5%0.7381-year OS45.4%0.694950.3%0.7563-year OS34.1%0.711739.9%0.709 Open table in a new tab

Females only account for a minority of malignant pleural mesothelioma (MPM) diagnoses, but may experience differential survival relative to men. It is unclear if differential receipt of treatment may be contributory. In this National Cancer Database analysis, we found that although surgery and chemotherapy are disproportionately underutilized in female patients with MPM, female gender is independently associated with improved survival relative to males. Background: Despite accounting for a minority of malignant pleural mesothelioma (MPM) diagnoses, females may experience differential survival relative to males. It is unclear if there are gender-based differences in receipt of treatment or disease-related outcomes for patients with MPM. We therefore utilized the National Cancer Database (NCDB) to assess patterns-of-care and overall survival (OS) among patients with MPM by gender. Materials and Methods: Patients with histologically confirmed MPM treated from 2004 to 2013 were identified from the NCDB. The association between female gender and OS was assessed using multivariable Cox proportional hazards models with propensity score matching. Patterns-of-care were assessed using multivariable logistic regression. The overall treatment effect was tested in subsets of patients by treatment strategy, histology, and clinical stage. Results: A total of 18,799 patients were identified, of whom 14,728 (78%) were male and 4071 (22%) were female. Females were statistically more likely to present at a younger age, with fewer comorbidities, and with epithelioid histology. Despite these favorable prognostic features, women were less likely to receive surgery (P <= .001) or chemotherapy (P <= .001) compared with males. On multivariable analysis, female gender was associated with improved OS (hazard ratio, 0.83; 95% confidence interval, 0.80-0.86; P <= .001). Gender-based survival differences were seen across all stages, but only among patients with epithelioid (P <= .001) and not biphasic (P = .17) or sarcomatoid (P = 1.00) histology. Conclusions: Surgery and chemotherapy are disproportionately underutilized in female patients with MPM. Despite this concerning disparity, female gender is independently associated with improved survival relative to males. Further research to understand factors that lead to gender disparities in MPM is warranted. (C) 2020 Elsevier Inc. All rights reserved.

We hypothesized that employing a machine learning approach could permit accurate prediction of unplanned hospitalizations, feeding tube placement, and significant weight loss experienced by head and neck (HN) cancer patients secondary to radiation therapy (RT). To test this, we merged data from an internal web-based charting tool (known as Brocade), the electronic health record (Epic), and the record/verify system to develop predictive models of these toxicities.

271 Background: Unplanned hospitalizations may diminish quality of care among cancer patients receiving radiotherapy (RT). In patients undergoing RT for gastrointestinal (GI) cancers, we hypothesized that a machine learning approach would enable prediction of unplanned hospitalizations within 30 days of RT. Methods: We analyzed 836 abdominal (gastric, pancreatic, biliary, hepatic) and 514 pelvic (rectal, anal) courses of RT for GI cancers treated at our institution (3/2016—1/2019). Over 700 clinical/treatment variables and unplanned hospitalizations during or within 30 days after RT were mined from institutional databases. Using machine learning, we developed random forest (RF), gradient boosted decision trees (XGB), and logistic models for unplanned hospitalizations. Models were trained on 670 abdominal and 423 pelvic cases. Five-fold cross-validation (CV) was used to select model type and hyperparameters, using area under the ROC curve (AUC) to measure performance. The best model was validated on the subsequent 166 abdominal and 91 pelvic cases. AUC>0.70 was deemed clinically valid. Results: Among 1,350 cases, incidence of 30-day unplanned hospitalization was 12.3% (13.3% abdominal cohort; 10.7% pelvic cohort). Model CV AUCs are shown in table. The best models were XGB and RF for the abdominal and pelvic cohorts, respectively. Their validation testing AUCs are shown in table. For all models tested, lab values (e.g. potassium, lipoproteins, hemoglobin) prior to RT were significant predictors of unplanned hospitalizations. In the abdominal cohort, pancreatic primary and total RT dose were important. For the pelvic cohort, body mass index was important. Median healthcare costs from RT start - 30 days post-RT were $69,108 in non-hospitalized patients and $119,844 in hospitalized patients. Conclusions: In GI cancer patients undergoing RT, a machine learning model identified patients at risk of 30-day unplanned hospitalization. Predictive analytics may be a key tool to help providers identify high-risk patients and optimize interventions, while improving quality and value of care. [Table: see text]

PURPOSE:The purpose of this study was to compare the effectiveness of ensemble methods (e.g., random forests) and single-model methods (e.g., logistic regression and decision trees) in predictive modeling of post-RT treatment failure and adverse events (AEs) for breast cancer patients using automatically extracted EMR data.METHODS:Data from 1967 consecutive breast radiotherapy (RT) courses at one institution between 2008 and 2015 were automatically extracted from EMRs and oncology information systems using extraction software. Over 230 variables were extracted spanning the following variable segments: patient demographics, medical/surgical history, tumor characteristics, RT treatment history, and AEs tracked using CTCAEv4.0. Treatment failure was extracted algorithmically by searching posttreatment encounters for evidence of local, nodal, or distant failure. Individual models were trained using decision trees, logistic regression, random forests, and boosted decision trees to predict treatment failures and AEs. Models were fit on 75% of the data and evaluated for probability calibration and area under the ROC curve (AUC) on the remaining test set. The impact of each variable segment was assessed by retraining without the segment and measuring change in AUC (ΔAUC).RESULTS:All AUC values were statistically significant (P < 0.05). Ensemble methods outperformed single-model methods across all outcomes. The best ensemble method outperformed decision trees and logistic regression by an average AUC of 0.053 and 0.034, respectively. Model probabilities were well calibrated as evidenced by calibration curves. Excluding the patient medical history variable segment led to the largest AUC reduction in all models (Average ΔAUC = -0.025), followed by RT treatment history (-0.021) and tumor information (-0.015).CONCLUSION:In this largest such study in breast cancer performed to date, automatically extracted EMR data provided a basis for reliable outcome predictions across multiple statistical methods. Ensemble methods provided substantial advantages over single-model methods. Patient medical history contributed the most to prediction quality.

Objectives: There are 2 main treatment paradigms recognized by the National Comprehensive Cancer Network for resectable malignant pleural mesothelioma (MPM): induction chemotherapy followed by resection (IC/R), and up-front resection with postoperative chemotherapy (R/PC). These paradigms are being compared in an accruing randomized phase II trial. In the absence of such completed trials, in this study we evaluated overall survival (OS) and postoperative outcomes of IC/R and R/PC. Methods: The National Cancer Database was queried for newly diagnosed epithelioid/biphasic MPM. Metastatic, node-positive, and/or cT4 disease was excluded, along with nondefinitive surgery and lack of chemotherapy. Multivariable logistic regression ascertained factors independently associated with induction chemotherapy delivery. Kaplan-Meier analysis was used to evaluate OS between cohorts; multivariable Cox proportional hazards modeling was used to assess factors associated with OS. Survival was also evaluated between propensity-matched populations. Last, postoperative outcomes were assessed between groups. Results: Overall, 361 patients (182 IC/R, 179 R/PC) were analyzed. Temporal trends revealed that IC/R is decreasing over time. Survival of the IC/R cohort was similar to that of R/PC patients (20.9 vs 21.7 months; P = .500); this persisted after propensity matching (20.8 vs 22.0 months; P = .270). However, patients who underwent IC/R experienced longer postoperative hospitalization (median 7 days vs 6 days; P = .001) and higher 30-day mortality (3.3% vs 0%; P = .020). Conclusions: To our knowledge, this is the only comparative investigation of the 2 major management paradigms of operable MPM. IC/R regimens are decreasing over time in the United States. Although associated with survival similar to R/PC, IC/R might be associated with worse postoperative outcomes. Careful induction chemotherapy patient selection is thus highly recommended.

This large investigation evaluates surgical practice patterns and survival by histology for malignant pleural mesothelioma. Histology independently affects survival. Surgery was associated with increased survival in patients with epithelioid and biphasic, but not sarcomatoid, disease. Introduction: For the 3 histologic subtypes of malignant pleural mesothelioma (MPM)-epithelioid, sarcomatoid, and biphasic-the magnitude of benefit with surgical management remains underdefined. Materials and Methods: The National Cancer Data Base was queried for newly diagnosed nonmetastatic MPM with known histology. Patients in each histologic group were dichotomized into those receiving gross macroscopic resection versus lack thereof/no surgery. Kaplan-Meier analysis evaluated overall survival (OS) between cohorts; multivariable Cox proportional hazards modeling assessed factors associated with OS. After propensity matching, survival was evaluated for each histologic subtype with and without surgery. Results: Overall, 4207 patients (68% epithelioid, 18% sarcomatoid, 13% biphasic) met the study criteria. Before propensity matching, patients with epithelioid disease experienced the highest median OS (14.4 months), followed by biphasic (9.5 months) and sarcomatoid (5.3 months) disease; this also persisted after propensity matching (P < .001). After propensity matching, surgery was associated with significantly improved OS for epithelioid (20.9 vs. 14.7 months, P < .001) and biphasic (14.5 vs. 8.8 months, P = .013) but not sarcomatoid (11.2 vs. 6.5 months, P = .140) disease. On multivariable analysis, factors predictive of poorer OS included advanced age, male gender, uninsured status, urban residence, treatment at community centers, and T4/N2 disease (all P < .05). Chemotherapy and surgery were independently associated with improved OS, as was histology (all P < .001). Conclusion: This large investigation evaluated surgical practice patterns and survival by histology for MPM and found that histology independently affects survival. Gross macroscopic resection is associated with significantly increased survival in epithelioid and biphasic, but not sarcomatoid, disease. However, the decision to perform surgery should continue to be individualized in light of available randomized data. (C) 2018 Elsevier Inc. All rights reserved.

As an engineering material, DNA is well suited for the construction of biochemical circuits and systems, because it is simple enough that its interactions can be rationally designed using Watson–Crick base pairing rules, yet the design space is remarkably rich. When designing DNA systems, this simplicity permits using functional sections of each strand, called domains, without considering particular nucleotide sequences. However, the actual sequences used may have interactions not predicted at the domain-level abstraction, and new rigorous analysis techniques are needed to determine the extent to which the chosen sequences conform to the system’s domain-level description. We have developed a computational method for verifying sequence-level systems by identifying discrepancies between the domain-level and sequence-level behaviour. This method takes a DNA system, as specified using the domain-level tool Peppercorn, and analyses data from the stochastic sequence-level simulator Multistrand and sequence-level thermodynamic analysis tool NUPACK to estimate important aspects of the system, such as reaction rate constants and secondary structure formation. These techniques, implemented as the Python package KinDA, will allow researchers to predict the kinetic and thermodynamic behaviour of domain-level systems after sequence assignment, as well as to detect violations of the intended behaviour.

Purpose: This study of a large, contemporary national database evaluated postoperative outcomes and overall survival (OS) for malignant pleural mesothelioma (MPM) by facility volume. Methods: The National Cancer Database was queried for newly-diagnosed non-metastatic MPM undergoing definitive surgery (extrapleural pneumonectomy (EPP) or pleurectomy/decortication (P/D)). Patients were dichotomized into those receiving therapy at a high-volume facility (HVF), defined a priori at the 90th percentile of case volume, with all others categorized as lower-volume facilities (LVFs). Statistics included multivariable logistic regression, Kaplan-Meier analysis, propensity-matching, and multivariable Cox proportional hazards modeling. Sensitivity analysis varied the dichotomized HVF-LVF cutoff and evaluated effects on postoperative outcomes and OS. Results: Of 1307 patients, 621 (48%) were treated at LVFs and 686 (52%) at HVFs. HVFs were more often in the Middle/South Atlantic regions, and less likely in New England, South, and Midwest. Notably, 75% of procedures at HVFs were P/Ds, versus 84% at LVFs (p < 0.001). Patients treated at HVFs experienced shorter length of postoperative hospitalization (p = 0.035), lower 30-day readmission rates (4.6% vs. 6.1%, p = 0.021), and lower 90-day mortality rates (10.0% vs. 14.6%, p = 0.029). Median OS for respective groups were 18 versus 15 months (p = 0.010), which were not significant following propensity-matching (p = 0.540). On multivariable analysis, facility volume did not independently predict for OS. Sensitivity analyses confirmed the postoperative outcomes and OS findings. Conclusions: This is the largest investigation to date assessing facility volume and outcomes following surgery for MPM. Although no independent effects on OS were observed, postoperative outcomes were more favorable at HVFs. These findings have implications for postoperative management, patient counseling, referring providers, and cost-effectiveness.

Precision oncology, an innovative paradigm employing big data and predictive analytics to predict outcomes and toxicity of cancer treatment, requires a large repository of data sources, which typically reside across multiple, siloed software platforms. Our institution has developed and implemented a web-based charting tool that collects clinician-entered and validated structured data including acute RT toxicities. We merged data from this tool (known as Brocade), the institutional EHR (Epic), the treatment planning system (Pinnacle), and the record and verify system (Mosaiq) to develop predictive models of acute toxicity during RT for breast cancer patients. From 05/2016—10/2017, 2,277 consecutive RT courses for breast cancer were administered across 5 practice sites within our Institution. For each course, >230 clinical and treatment variables were collected from the aforementioned information systems. Acute toxicity outcomes included moist desquamation and NCI CTCAEv4 grade ≥2 radiation dermatitis, breast/chest wall pain, and fatigue; all outcomes were documented by the treating physician in a structured format during RT using Brocade. Random forest, gradient boosted decision tree, and logistic regression models were trained on the initial 1,977 RT courses to predict occurrence for each outcome. Five-fold cross-validation (CV) was used to select model type and hyperparameters, using area under the ROC curve (AUC) to measure performance. The best performing model for each outcome was then evaluated on an independent validation set consisting of the subsequent 300 consecutive courses of RT for breast cancer. Models with an AUC > 0.70 were considered clinically valid. Among the 2,277 patients, 99.6% were female and median age was 58 years (IQR: 48-66). The average AUCs across CV folds of the training set for all outcomes are listed in the table, as well as incidence of acute toxicities. All AUCs in the training set were >0.70. In the validation set, the incidence of radiation dermatitis, moist desquamation, and breast/chest wall pain was 27.7%, 6.7%, and 2.7%, respectively. The AUC values for these toxicity models were 0.85, 0.82, 0.77, respectively, meeting the threshold for clinical validity. Fatigue, noted in 3.7% of patients, had an AUC of 0.56. Application of this machine learning-based approach yielded clinically valid models for moist desquamation, grade ≥2 radiation dermatitis, and grade ≥2 breast pain. To our knowledge, this is the first demonstration of the ability of a precision oncology approach to accurately predict acute RT toxicities in a prospective validation dataset. This approach could help identify patients who may benefit from early interventions to avert acute RT toxicity.Abstract 116; TableTraining Set (n=1977)Dermatitis (35.6%)Moist desquamation (9.5%)Breast pain (3.8%)Fatigue (2.6%)Random forest0.8070.8080.6960.681Gradient boosted decision trees0.8110.8080.6770.704Logistic regression0.8130.8120.6910.672 Open table in a new tab

Introduction: Controversy exists regarding the optimal surgical technique for malignant pleural mesothelioma (MPM). We evaluated national practice patterns and outcomes of MPM treated with extrapleural pneumonectomy (EPP) versus lung-sparing extended pleurectomy/decortication (P/D). Methods: The National Cancer Database was queried for patients with newly diagnosed MPM undergoing EPP or P/D. Multivariable logistic regression ascertained clinical factors independently associated with P/D receipt. Kaplan-Meier analysis was used to evaluate overall survival (OS) between cohorts; multivariable Cox proportional hazards modeling was used to evaluate factors associated with OS. Survival was then evaluated between propensity-matched populations. Results: Overall, 1307 patients (271 undergoing EPP [21%] and 1036 undergoing P/D [79%]) met the criteria. Patients receiving P/D were older (p = 0.028), whereas those undergoing EPP were more likely to live in a rural area (p = 0.044), live farther from the treating facility (p = 0.039), and receive treatment at an academic center (p = 0.050). There were no differences between cohorts in 30-day readmission or mortality (all p > 0.05). The median OS times in the EPP and P/D groups were 19 versus 16 months, respectively (p = 0.120); no differences were observed after propensity matching (p = 0.540). Conclusions: In this largest analysis of its kind to date, findings from this contemporary cohort demonstrate that P/D comprised most surgical procedures for MPM. Procedure type was influenced by sociodemographic and geographical factors, without observed differences in survival or postoperative mortality and readmission rates between techniques. (C) 2017 International Association for the Study of Lung Cancer. Published by Elsevier Inc. All rights reserved.

One form of precision medicine in radiation oncology involves guiding treatment decisions based on personalized predictions of clinical outcomes. For this to be impactful, high-quality predictive models must be built, and the use of outcome predictions must affect clinical practice. We aimed to quantify the potential for predictive models to impact clinical practice for different patient groups, decoupling the impact of predictive models from their construction. We hypothesize that high potential for impact exists in lung and head and neck cancers due to high patient volume, patient and tumor variability, and high toxicity rates. In this IRB-approved analysis, a dataset of 16,689 radiotherapy courses performed at one institution between 2008 and 2015 was created from EMRs and treatment planning systems. We extracted 295 variables spanning demographics, tumor characteristics, medical, surgical, and radiation history, and adverse events (scored using CTCAEv4.0). We defined "potential impact" as potential to personalize treatments in ways that improve outcomes for large numbers of patients. For 14 patient groups, we computed four separate measures: potential for personalization (joint entropy of patient variables, H(P), and treatment variables, H(T)), potential for outcome improvement (joint entropy of outcome variables, H(O), and average number of distinct adverse events per patient (AAE)), potential for treatment-outcome link (conditional mutual information between outcomes and treatments given patient variables, I(O;T|P)), and potential for high-volume impact (group size as a fraction of all patients). To assess relative differences in patient groups, we ranked each disease site within each measure. The top three groups for patient volume and personalization score were head and neck, lung, and brain. Head and neck, lung, and lower gastrointestinal malignancies were the top groups in outcome improvement potential. Prostate, breast, and brain were the top three for treatment-outcome link. The top five groups overall, measured by the sum of the four measure-specific ranks, are summarized in the tableAbstract 3002; Table 1Quantifying the Potential Impact of Predictive Analytics in Radiation OncologyVolumePersonalizationImprovementT-O LinkOverallCasesFractionH(P)H(T)H(O)AAEI(O;T|P)Sum of ranksLung20600.12310.953.419.546.10.0411Head&Neck21300.12711.043.479.246.50.00812Brain20090.1210.863.499.045.20.0514Lower GI5000.038.913.567.676.80.0419Prostate17000.10210.182.169.85.10.2520 Open table in a new tab Our study, using a novel method to quantify potential impact of predictive analytics in radiation oncology, showed the greatest potential for personalized therapies in lung, head and neck, and brain malignancies, likely due to high rates of adverse events, tumor variability, and high patient numbers. Other malignancies, like prostate cancer, may be easier to build powerful predictive models due to a high score for treatment-outcome link. Patients from additional centers are needed to validate our findings. Clinical trials incorporating predictive analytics are planned.