Search

Background The aim of this study is to provide a comprehensive understanding of the current landscape of artificial intelligence (AI) for cancer clinical trial enrollment and its predictive accuracy in identifying eligible patients for inclusion in such trials. Methods Databases of PubMed, Embase, and Cochrane CENTRAL were searched until June 2022. Articles were included if they reported on AI actively being used in the clinical trial enrollment process. Narrative synthesis was conducted among all extracted data: accuracy, sensitivity, specificity, positive predictive value, and negative predictive value. For studies where the 2x2 contingency table could be calculated or supplied by authors, a meta-analysis to calculate summary statistics was conducted using the hierarchical summary receiver operating characteristics curve model. Results Ten articles reporting on more than 50 000 patients in 19 datasets were included. Accuracy, sensitivity, and specificity exceeded 80% in all but 1 dataset. Positive predictive value exceeded 80% in 5 of 17 datasets. Negative predictive value exceeded 80% in all datasets. Summary sensitivity was 90.5% (95% confidence interval [CI] = 70.9% to 97.4%); summary specificity was 99.3% (95% CI = 81.8% to 99.9%). Conclusions AI demonstrated comparable, if not superior, performance to manual screening for patient enrollment into cancer clinical trials. As well, AI is highly efficient, requiring less time and human resources to screen patients. AI should be further investigated and implemented for patient recruitment into cancer clinical trials. Future research should validate the use of AI for clinical trials enrollment in less resource-rich regions and ensure broad inclusion for generalizability to all sexes, ages, and ethnicities.

Purpose Higher energy (>6 MV) photons reduce dose inhomogeneity with breast tangent beams, thereby reducing late breast toxicity, but skin and superficial tissue sparing by higher energy beams raises concerns about local recurrence (LR) risk. This study aimed to determine whether beam energy and surgical bed-to-skin distance affect LR. Methods and Materials This population-based study included newly diagnosed invasive breast cancers without skin involvement (pT1-4a, any-N, M0) treated with breast-conserving surgery and adjuvant whole breast radiotherapy without bolus or beam spoilers. The primary endpoint was the cumulative incidence of LR (CILR). A multivariable analysis (MVA) included mean beam energy, age, T-stage, nodal status, overall stage, lymphovascular invasion (LVI), grade, margin status, extensive intraductal component (EIC), breast cancer subtype, hormone therapy and chemotherapy. In a subgroup with contoured surgical beds, another MVA included surgical bed-to-skin distance. Results The cohort consisted of 10,083 women treated from 2002 to 2011, 327 with 4MV, 6,006 with 6 MV, 2,083 with >6-10 MV and 1,667 with >10 MV tangents. The median follow-up time was 11.1 years. The 10-year CILR was 3.1% [95% confidence interval 1.6,5.4] with 4 MV, 2.8% [2.4,3.3] with 6 MV, 4.2% [3.4,5.3] with >6-10 MV and 2.6% [1.9,3.5] with >10 MV. On MVA of the entire cohort, LR risk was increased with positive margins, LVI, EIC, and lack of hormone therapy, but was not associated with beam energy (HR = 1.01 [0.96,1.05], p = 0.8). On MVA of 3,359 patients with contoured surgical beds, LR risk was not associated with surgical bed-to-skin distance (HR = 1.00 [0.99,1.02], p = 0.8). Conclusions Use of higher breast tangent beam energies is not associated with increased risk of local recurrence, including in cases with surgical beds that are close to the skin.

Background: The timely delivery of radiotherapy (RT) is crucial to cancer care, and excessive delays in RT have been associated with detrimental oncological and psychological outcomes. Prior to receiving RT on treatment units (Linear accelerators), there are a few processes that need to take place including simulation (on CT simulators), radiotherapy plan generation/optimization and quality assurance. The assignment of patient schedules on CT simulators and Linear accelerators is currently done manually at most cancer centers. We propose that data-driven optimization of patient scheduling has the potential to improve wait-times, and optimize use of departmental resources. Methods: A two-stage Mixed Integer Programming model was developed to optimize the patient appointment scheduling process and to forecast machine utilization. The model was tested with historical institutional data from Princess Margaret Cancer Center. By analyzing the dataset and simulating historical patient arrivals, the model output is evaluated by comparing patient wait time statistics and monthly machine utilization against what occurred during this time frame. Results: Testing our model on data from 2019-06 to 2020-02, we found a reduction in average wait time from 11.2 to 6.7 business days for standard category patients. The number of standard patients exceeding the wait time target of 10 business days were reduced from 118 to 15 patients each month. In addition, our model could accurately estimate future machine utilization for both CT simulators and linear accelerators based on the model output appointments, which could facilitate better planning and utilization of departmental resources. Conclusion: Our scheduling model has the potential to reduce the standard patient wait time for radiation treatment without compromising the wait time for urgent patients. The model can be also used to forecast department resources and machine utilization based on the output of the scheduling model. Radiotherapy departments could use this model to generate patient appointment schedules as well as to reduce machine idle time or appointment over-booking.

PURPOSE Clinical TNM staging is a key prognostic factor for patients with lung cancer and is used to inform treatment and monitoring. Computed tomography (CT) plays a central role in defining the stage of disease. Deep learning applied to pretreatment CTs may offer additional, individualized prognostic information to facilitate more precise mortality risk prediction and stratification. METHODS We developed a fully automated imaging-based prognostication technique (IPRO) using deep learning to predict 1-year, 2-year, and 5-year mortality from pretreatment CTs of patients with stage I-IV lung cancer. Using six publicly available data sets from The Cancer Imaging Archive, we performed a retrospective five-fold cross-validation using pretreatment CTs of 1,689 patients, of whom 1,110 were diagnosed with non–small-cell lung cancer and had available TNM staging information. We compared the association of IPRO and TNM staging with patients' survival status and assessed an Ensemble risk score that combines IPRO and TNM staging. Finally, we evaluated IPRO's ability to stratify patients within TNM stages using hazard ratios (HRs) and Kaplan-Meier curves. RESULTS IPRO showed similar prognostic power (concordance index [C-index] 1-year: 0.72, 2-year: 0.70, 5-year: 0.68) compared with that of TNM staging (C-index 1-year: 0.71, 2-year: 0.71, 5-year: 0.70) in predicting 1-year, 2-year, and 5-year mortality. The Ensemble risk score yielded superior performance across all time points (C-index 1-year: 0.77, 2-year: 0.77, 5-year: 0.76). IPRO stratified patients within TNM stages, discriminating between highest- and lowest-risk quintiles in stages I (HR: 8.60), II (HR: 5.03), III (HR: 3.18), and IV (HR: 1.91). CONCLUSION Deep learning applied to pretreatment CT combined with TNM staging enhances prognostication and risk stratification in patients with lung cancer.

Background The hypothesis of a curable oligometastatic prostate cancer (PCa) state remains to be clinically-proven. Conventional imaging often fails to localize early recurrences, hampering the potential for radical approaches. Objective We hypothesize that prostate-specific membrane antigen (PSMA)-targeted PET-MR/CT allows for earlier detection and localization of oligorecurrent-PCa, unveiling a molecularly-defined state amenable to curative-intent metastasis-directed treatment (MDT). Design/setting/participants Single-institution single-arm phase-two study. Patients with rising PSA (0.4-3.0 ng/mL) after maximal local therapy (radical prostatectomy and post-operative radiotherapy), negative conventional staging, and no prior salvage hormonal therapy (HT) were eligible. Interventions All patients underwent [18F]DCFPyL PET-MR/CT. Patients with molecularly-defined oligorecurrent-PCa had MDT (stereotactic ablative body radiotherapy [SABR] or surgery) without HT. Outcome measurements/statistical analysis Primary endpoint was biochemical response (complete, i.e. biochemical ‘no evidence of disease’ [bNED], or partial response [100% or ≥50% PSA decline from baseline, respectively]) after MDT. Simon’s two-stage design was employed (null and alternate hypotheses 20% response rate, respectively), with α and β of 0.1. Results Seventy-two patients were enrolled (May/2017-July/2019). Thirty-eight (53%) had PSMA-detected oligorecurrent-PCa amenable for MDT. Thirty-seven (51%) agreed to MDT: 10 and 27 underwent surgery and SABR, respectively. Median follow-up was 15.9 months (IQR 9.8-19.1). Of patients receiving MDT, the overall response rate was 60%, including 22% rendered bNED. One (2.7%) grade 3 toxicity (intra-operative ureteric injury) was observed. Conclusions PSMA-defined oligorecurrent-PCa can be rendered bNED, a necessary step towards cure, in 1 of 5 patients receiving MDT alone. Randomized trials are justified to determine if MDT +/− systemic agents can expand the curative therapeutic armamentarium for PCa. Patient summary We studied men treated for prostate cancer with rising PSA. We found PSMA imaging detected recurrent cancer in three-quarters of patients, and targeted treatment to these areas significantly decreased PSA in half of patients.

Highlights • Prophylactic cranial irradiation (PCI) was more commonly used in younger patients. • PCI utilization rates did not change throughout our 20-year institutional experience. • PCI was associated with improved OS and lower brain metastasis risk, independent of MRI follow-up or era of treatment. • For LS-SCLC patients with good thoracic response, PCI remains the standard-of-care., Background & purpose Prophylactic cranial irradiation (PCI) is recommended for limited-stage small-cell lung cancer (LS-SCLC) patients with good response to concurrent chemoradiation. We report our institution’s 20-year experience with this patient population and associated clinical outcomes. Materials & methods A retrospective cohort of consecutive LS-SCLC patients treated with curative intent chemoradiation at our institution (1997–2018) was reviewed. Overall survival (OS) was calculated using the Kaplan-Meier method, and significant covariates determined by the Cox proportional hazards model. Covariates predictive of PCI were determined using Fisher's exact test and the Mann-Whitney test. Brain failure risk (BFR) was calculated using the cumulative incidence method treating death as a competing event. Treatment cohorts (historic vs. contemporary) were stratified by the median year of diagnosis (2005). Results A total of 369 patients with LS-SCLC were identified, of which 278 patients were notionally PCI eligible. PCI was given to 196 patients (71%). Younger age was associated with PCI utilization (p

Purpose Bolus use during postmastectomy radiation therapy doubles the risk of grade 2 and 3 skin toxicity. Despite its unknown benefit, bolus is often prescribed during postmastectomy radiation therapy for patients without skin involvement. Methods and Materials For women with breast cancer receiving photon 3-dimensional conformal radiation therapy, bolus was used routinely for chest walls but was omitted for breast reconstructions by about half of radiation oncologists from 2007 to 2011. Eligible patients had newly diagnosed invasive breast cancers without skin involvement (pT1-4a, any-N, M0) treated with adjuvant or neoadjuvant radiation therapy. For the bolus and no-bolus groups, we compared the cumulative incidence of local recurrence (LR) and locoregional recurrence (LRR) with distant recurrence and death as competing risks and breast cancer mortality (BCM). Multivariable analysis of LR and BCM included stage, subtype, lymphovascular invasion, grade, margin status, beam energy, bolus use, hormone therapy, chemotherapy, and reconstruction. Results Systemic therapy was used for 98% of the 1887 patients. The bolus group had 1569 patients and the no-bolus group had 318 patients. Bolus was used in 51% (281/550) of patients treated with reconstruction and 96% (1288/1337) of patients treated without reconstruction. The 10-year outcomes (95% confidence interval) in patients treated with and without bolus were, respectively: LR 1.9% (1.3-2.7) versus 0.9% (0.3-2.6), LRR 3.1% (2.3-4.0) versus 3.2% (1.6-5.5), and BCM 19.4% (17.3-21.6) versus 18.3% (13.9-23.2). On multivariable analysis, bolus use was not associated with better LR (hazard ratio = 1.4 [0.3-6.4]) or BCM (hazard ratio = 0.8 [0.5-1.2]). Conclusions For patients treated with mastectomy, radiation therapy, and modern systemic therapy, the cumulative incidence of LR was low, with or without bolus. Because bolus use increases toxicity and does not reduce LR or BCM, it should no longer be used routinely for patients without skin involvement who receive systemic therapy.

Machine learning (ML) holds great promise for impacting healthcare delivery; however, to date most methods are tested in ‘simulated’ environments that cannot recapitulate factors influencing real-world clinical practice. We prospectively deployed and evaluated a random forest algorithm for therapeutic curative-intent radiation therapy (RT) treatment planning for prostate cancer in a blinded, head-to-head study with full integration into the clinical workflow. ML- and human-generated RT treatment plans were directly compared in a retrospective simulation with retesting (n = 50) and a prospective clinical deployment (n = 50) phase. Consistently throughout the study phases, treating physicians assessed ML- and human-generated RT treatment plans in a blinded manner following a priori defined standardized criteria and peer review processes, with the selected RT plan in the prospective phase delivered for patient treatment. Overall, 89% of ML-generated RT plans were considered clinically acceptable and 72% were selected over human-generated RT plans in head-to-head comparisons. RT planning using ML reduced the median time required for the entire RT planning process by 60.1% (118 to 47 h). While ML RT plan acceptability remained stable between the simulation and deployment phases (92 versus 86%), the number of ML RT plans selected for treatment was significantly reduced (83 versus 61%, respectively). These findings highlight that retrospective or simulated evaluation of ML methods, even under expert blinded review, may not be representative of algorithm acceptance in a real-world clinical setting when patient care is at stake. An artificial intelligence system prospectively deployed to design radiation therapy plans for patients with prostate cancer illustrates the real-world impact of machine learning in clinical practice and identifies factors influencing human–algorithm interaction

To compare the long-term oncologic outcomes of intermediate risk (IR) prostate cancer (PCa) patients treated with low dose-rate brachytherapy (LDR-BT) or moderate hypofractionated external beam radiotherapy (HF-EBRT).Patients diagnosed with IR PCa and treated with LDR-BT or HF-EBRT between January 2005 and December 2013 were included. Brachytherapy treatment involved a transperineal implant of iodine-125 to a dose of 145 Gy to the PTV, while HF-EBRT was delivered using intensity modulated radiotherapy with 60 Gy in 20 fractions. The Phoenix ''nadir +2'' threshold was used to define biochemical relapse (BR). The cumulative incidence function (CIF) of BR and metastases was reported for each group and compared using the Gray's test to account for the competing risk of death. The Kaplan-Meier (KM) method was used to estimate overall survival (OS) and prostate cancer specific survival (PCSS). Univariate (UVA) and multivariable (MVA) analysis of the CIF of BR and metastases were performed. A 2-tailed p-value ≤ 0.05 was considered statistically significant.Overall, 122 and 124 patients were treated with LDR-BT and HF-EBRT respectively. Median follow-up was 95 months [interquartile range (IQR): 79-118] in the LDR-BT group and 96 months (IQR: 63-123) in the HF-EBRT group. BR was observed in 5 patients treated with LDR-BT and 34 treated with HF-EBRT. At 60 and 90 months, the CIF of BR was 0.9% and 3.5% in the LDR-BT group vs. 16.6% and 23.7% in the HF-EBRT (p0.001). The CIF of metastases at 90 and 108 months, was 0% and 1.6% vs. 3.4% and 9.1% in the LDR-BT and HF-EBRT groups (p = 0.003), respectively. At the last follow-up, 3 patients treated with HF-EBRT died from their cancer [PCSS of 97.5% at 8 years and none died in the LDR-BT group (p = 0.09). On UVA and MVA risk group and treatment modality were independently associated with CIF of BR. On UVA HF-EBRT and ISUP grade group 3 were associated with metastases.LDR-BT was associated with higher biochemical and metastases control in our cohort when compared to moderately HF-EBRT. In the absence of a randomized trial, LDR-BT when feasible should be offered to patients with a life expectancy of8 years.

Introduction: The addition of durvalumab after chemoradiation therapy (CRT) in unresectable stage III NSCLC significantly improves survival. The benefit of this approach in elderly patients is controversial given the toxicity associated with CRT and, thus, may be underutilized. We sought to investigate the outcomes of elderly patients treated with CRT without or without durvalumab at our center. Methods: We reviewed all stage III patients with NSCLC treated with CRT between 2018 and 2020. Patients were analyzed on the basis of age: less than 70 years and 70 years and older. The end points evaluated were treatment patterns, toxicity, progression-free survival, and overall survival. Results: The baseline characteristics including Eastern Cooperative Oncology Group performance status and comorbidities were similar among the 115 patients (44 elderly, 71 young). Completion rates of CRT (100%, 97%) and chemotherapy dose intensity (97%, 97%) were high in elderly and young patients, respectively. There was a trend toward increased hospitalizations in elderly patients because of infections (27% versus 13%, p = 0.08). Of those who did not have primary progression after CRT, 78% of eldery and 81% of young patients received durvalumab. The incidence of grade 3 or higher immune-related adverse events was 9% in elderly and 6% in young patients (p = 0.67). The median progression-free survival was similar (15.6 versus 10.5 mo, p = 0.10), even after adjusting for comorbidities (hazard ratio = 0.6, p = 0.09). The 12-month overall survival rates were 78% in the elderly and 76% in young patients (p = 0.98). Conclusions: Well-selected elderly patients can be treated safely with CRT followed by durvalumab with similar survival benefits compared with their younger counterparts. We would advocate for the referral of all elderly patients for oncologic assessment to avoid undertreatment.

Urine cytology and cystoscopy are routinely employed during follow-up of patients after trimodal therapy (TMT) for muscle-invasive bladder cancer (MIBC). The significance of positive or equivocal cytology without visible disease recurrence on cystoscopy during follow-up is unknown, and studies informing outcomes in this scenario are lacking. This study aims to investigate the temporal trends of positive/equivocal cytology in the absence of visible disease recurrence and the association with bladder cancer recurrence and survival outcomes.One hundred and twenty-nine patients with available post-TMT cytology data and negative cystoscopy from a single academic institution between 2002 and 2017 with a median follow-up of 3.4 (range 0.1-14.2) years were analyzed. Cytology results, first post-TMT cytology positive/equivocal (CP) and negative (CN), were evaluated for association with disease recurrence and survival. Kaplan. Meier and competing risks methods were used to assess time-to-negative cytology in CP patients with ≥2 interval post-TMT cytology results (n = 33), time-to-recurrence, and disease-specific mortality (DSM) stratified by first post-TMT cytology result.At first follow-up (6-8 weeks post-TMT completion), CP was observed in 41 (32%) and CN in 88 (68%) of patients. With further follow-up of CP patients with ≥2 interval post-TMT cytology results, the probability of developing negative cytology was 57% (95% CI 42, 77) at 6 months post-TMT, and the median time-to-negative cytology was 3.2 months (95% CI 2.99, 5.80). The median time-to-recurrence was reduced in CP patients compared to CN (24.3 vs. 78.1 months, p = 0.1), corresponding with an apparent increase in the cumulative incidence of recurrence rate at 3 years in the CP vs. CN group (62% vs. 42%, p = 0.1). No significant difference was observed in the 3-year DSM rates. On univariable analysis, the hazards of recurrence and DSM for patients with CP were 1.5 (95% CI 0.9, 2.5, p = 0.1) and 2.1 (95% CI 0.9, 4.7, p = 0.07) respectively.This is the first study to investigate the significance of a positive/equivocal cytology without visible disease following TMT for MIBC. Positive cytology is common and does not preclude subsequent negative cytology supporting a watchful waiting approach rather than proceeding immediately to biopsy. However, cytology that remains positive at subsequent follow-up may be associated with adverse recurrence and survival outcomes.

The aim of this study was to characterize the patterns of prostate cancer recurrence after brachytherapy (BT) using 2-(3-[1-carboxy-5-([6-Patients were selected from an ongoing prospective institutional trial investigating the use of [Between March 2017 and April 2020, 670 patients underwent [Contrary to previous evidence, our study suggests that in prostate BT patients with biochemical recurrence, the most common site of failure is local for the patients treated with monotherapy and metastatic for patients treated with a combination of external beam radiation and BT boost.

Purpose/Objective(s) Multidisciplinary tumor boards (MDTs) are an opportunity for specialists to meet and discuss the diagnosis and management of cancer patients. In our institution, patients with lung metastases are reviewed at a specialized tumor board comprising radiation oncologists, medical oncologists, surgeons, and radiologists. During these meetings, treatment recommendations for patients are made after a careful review of the patients' clinical history and relevant investigations. The objective of this study is to develop a machine learning classifier that reproduces MDT decision making and can be used as a decision support system to facilitate quality assurance and knowledge sharing in future MDTs. Materials/Methods Treatment decisions were manually gathered from patients discussed by the lung metastases MDT from 2016 to 2020 through clinical notes in the electronic medical records. Twenty-two variables were collected, including: age, gender, primary cancer histology, the number of lesions, the size of lesions, the presence of mediastinal/hilar lymphatic spread, the number of lobes/anatomical regions involved, disease free interval, ECOG status, and the presence of extra thoracic metastases. The target variable was binary: local therapy (SBRT or surgery) vs non-local therapy (systemic therapy, surveillance and palliative radiation). The dataset was split into a training, validation, and test set using a 60/20/20 split. Multiple machine learning models were evaluated (logistic regression, decision tree, support vector machine, naive Bayes, random forest) with a combination of feature engineering, feature selection, and hyper parameter tuning. Results The dataset contains treatment decisions for 506 patients. The ages of patients ranged from 17 to 92 years old, with 215 female and 291 male patients. The most common primary cancers included colorectal (105), sarcoma (102), and head and neck (80). The dataset consisted of 235 local therapy, 111 systemic therapy, 152 surveillance, and 8 palliative radiation decisions. The best performing model was a random forest classifier, which classified local therapy vs non-local therapy decisions with an AUC of 86.6% and accuracy of 80.2%. The two most predictive features were the number of lesions and number of lobes/anatomical regions involved with metastases. Conclusion This study demonstrates that relatively simple variables taken from patient records can be used to create a model that generates recommendations concordant with an MDT. The next steps of the study include extracting and processing cross-sectional imaging to acquire exact tumor dimensions, density and location with respect to the pleura, fissures, bronchi, and vessels. This may improve the current classifier performance, as well as allow for more detailed predictions such as further classifying local treatment decisions into surgery and SBRT.

PURPOSE/OBJECTIVE(S) Despite evidence for the superiority of twice daily (BID) radiotherapy schedules, their utilization in practice remains logistically challenging. Hypofractionation (HFRT) is a commonly implemented alternative. We aim to compare the outcomes and toxicities in limited stage small cell lung cancer (LS-SCLC) patients treated with hypofractionated versus BID schedules. MATERIALS/METHODS A bi-institutional retrospective cohort review was conducted of LS-SCLC patients treated with BID (45Gy/30 fractions) or HFRT (40Gy/15 fractions) schedules from 2007-2019. Overlap weighting using propensity scores were performed to balance observed covariates between the two radiotherapy schedule groups. Effect estimates of radiotherapy schedule on overall survival (OS), locoregional recurrence (LRR) risk, thoracic response, any grade 3+ (including lung, and esophageal) toxicity were determined using multivariable regression modelling. E-values were used to assess the sensitivity of effect estimate to unobserved confounding. RESULTS A total of 173 patients were included in the overlap weighted analysis, with 110 patients having received BID treatment, and 63 treated by HFRT. The 5-year OS for the overlap weighted cohort was 24.3% (95% CI, 16.1-36.6), and specifically 22.1% (95% CI, 12.7-38.5) and 26.6% (95% CI, 14.4-49.0%) when stratified by BID and HFRT cohorts specifically. The 5-year LRR risk for the same cohorts after overlap weighting were 68.9% (95% CI, 59.2-80.1), 68.9% (95% CI, 56.6-83.8), and 69.2% (95% CI, 55.3-86.6), respectively. Multivariable regression modelling did not reveal any significant differences in OS (hazard ratio [HR] 1.67, P = 0.38), LRR risk (HR 1.48, P = 0.38), CRT response (odds ratio [OR] 0.23, P = 0.21), any grade 3+ toxicity (OR 1.67, P = 0.33), grade 3+ pneumonitis (OR 1.14, P = 0.84), or grade 3+ esophagitis (OR 1.41, P = 0.62). These effect estimates were consistent between unweighted and overlap weighted cohorts. CONCLUSION HFRT, in comparison to BID radiotherapy schedules, does not appear to result in significantly different survival, locoregional control, or toxicity outcomes.

e20575 Background: Computed tomography (CT) imaging is used to inform staging and treatment decisions for stage I non-small cell lung cancer (NSCLC) patients. We have previously used deep learning applied to pretreatment CTs to generate an imaging-based prognostication (IPRO) score that automatically quantifies mortality risk and stratifies patients beyond tumor, node, metastasis (TNM) substages. Here we present validation data of its prognostic impact. Methods: We developed a fully automated deep learning model, IPRO, designed to process a CT scan, localize the 36cm3 space centered on the lungs, and learn prognostic imaging features to predict mortality risk. IPRO was trained on pretreatment CTs acquired from 1,696 patients treated for NSCLC at a tertiary care center between 2004 and 2018. We withheld 20% of the cases for validation, including 162 patients that were diagnosed with stage I NSCLC by clinical staging. We evaluated IPRO’s ability to stratify stage I NSCLC patients into mortality risk quintiles using the Cox proportional hazards model and assessed differences in median overall survival (mOS). Results: Of the 162 stage I NSCLC patients in the validation set, the mOS was 68.5 months (95% CI 66.7-69.6), 85 (52.5%) were male, and 125 (77.2%) were diagnosed with stage IA. Of these, 111 patients received surgery, 40 received radiotherapy (RT), 9 received surgery + adjuvant systemic therapy (ST), and 2 patients received surgery + ST + RT. According to IPRO, the patients predicted to have the highest risk had significantly increased 5-year mortality compared to those predicted to have the lowest risk (OR 8.4, 95% CI 2.4-29.2, p < 0.01); median survival 48.0 months and 69.5 months, respectively. Conclusions: Deep learning applied to pretreatment CTs provides personalized prognostic insights for stage I NSCLC beyond current TNM staging. [Table: see text]

Machine learning (ML) holds great promise for impacting healthcare delivery; however, to date most methods are tested in 'simulated' environments that cannot recapitulate factors influencing real-world clinical practice. We prospectively deployed and evaluated a random forest algorithm for therapeutic curative-intent radiation therapy (RT) treatment planning for prostate cancer in a blinded, head-to-head study with full integration into the clinical workflow. ML- and human-generated RT treatment plans were directly compared in a retrospective simulation with retesting (n = 50) and a prospective clinical deployment (n = 50) phase. Consistently throughout the study phases, treating physicians assessed ML- and human-generated RT treatment plans in a blinded manner following a priori defined standardized criteria and peer review processes, with the selected RT plan in the prospective phase delivered for patient treatment. Overall, 89% of ML-generated RT plans were considered clinically acceptable and 72% were selected over human-generated RT plans in head-to-head comparisons. RT planning using ML reduced the median time required for the entire RT planning process by 60.1% (118 to 47 h). While ML RT plan acceptability remained stable between the simulation and deployment phases (92 versus 86%), the number of ML RT plans selected for treatment was significantly reduced (83 versus 61%, respectively). These findings highlight that retrospective or simulated evaluation of ML methods, even under expert blinded review, may not be representative of algorithm acceptance in a real-world clinical setting when patient care is at stake.