Main Session
Sep
28
PQA 03 - Digital Health Innovation and Informatics, Patient Safety & Quality, and Radiation and Cancer Biology
Presenter(s)
Seyyedeh Azar Oliaei Motlagh, - University of Michigan, Ann Arbor, MI
S. A. Oliaei Motlagh1, C. Mayo2, B. S. Rosen3, K. C. Cuneo3, M. Yan4, T. Stanescu5, L. A. Dawson6, and T. S. Lawrence3; 1University of Michigan, Ann Arbor, MI, 2University of Michigan Medical School, Ann Arbor, MI, 3Department of Radiation Oncology, University of Michigan, Ann Arbor, MI, 4Radiation Medicine Program, Princess Margaret Cancer Centre, University Health Network, Toronto, ON, Canada, 5Medical Physics, Princess Margaret Cancer Centre, University Health Network, Toronto, ON, Canada, 6Department of Radiation Oncology, Princess Margaret Cancer Centre, University of Toronto, Toronto, ON, Canada
Purpose/Objective(s):
We developed a comprehensive pipeline linking big data repositories with automated statistical analysis and artificial intelligence (AI) modeling to construct a digital twin dashboard and metric system (DTDBM).Materials/Methods:
We previously demonstrated an automated approach of statistical profiling and AI modeling for distilling evidence from a high-dimensional multi-modal set of 436 features and thresholds for 290 hepatocellular carcinoma (HCC) patients treated with SBRT. Laboratory measures at 9 months for ALBI score, its change (?ALBI), and alkaline phosphatase (ALP) surpassed end of treatment (EOT) measures. Although dosimetric parameters were included in the analysis, laboratory values—particularly those obtained at 9 months—consistently demonstrated greater predictive power for overall survival (OS) compared to models based solely on EOT laboratory data. The optimal OS model using 9-month laboratory values achieved a diagnostic odds ratio (DOR) of 7.80 and an area under the curve (AUC) of 0.81. Building on these findings, we developed a DTDBM for personalized survival prediction, which was evaluated in an external validation cohort of 103 HCC patients treated with SBRT.Results:
From the external dataset, 75, 44, and 26 patients had follow-up data at least 1, 2, and 3 years after EOT, respectively, and were included in the corresponding digital twin predictions of OS. Among these, only 49, 30, and 20 patients had laboratory values available at 9 months. Using EOT laboratory values (ALBI and ALP), the mean absolute error (MAE) for OS prediction between the matched dataset and actual survival times of external patients was 7.4% at one year and 22.1% at two years. When 9-month laboratory values (ALBI, ALP, and ?ALBI) were utilized, prediction accuracy improved, with MAE decreasing to 2.9% at one year and 18.1% at two years. Statistical analyses confirmed the significance of these improvements, with paired t-test and Wilcoxon p-values less than 0.01. Survival curves for matched patient groups were comparable between institutions, as demonstrated by a non-significant log-rank test(test statistic = 0.09, p = 0.77).Conclusion:
Cross-institutional validation supported the digital twin model’s personalized survival prediction for HCC patients treated with SBRT, especially when using laboratory values at nine months post-treatment. These findings underscore the value of later follow-up laboratory data collection and support the integration of DTDBM into future clinical workflows and trials.