Patient-specific intervention analysis
in longitudinal glioblastoma
Patient-067's observed longitudinal imaging and surgical history, followed by chemoradiation, resection planning and comparative intervention analysis from her measured tumor states. Patients 071 and 072 illustrate matched clinical states with divergent observed courses. Each module records the session and produces a downloadable report.
Observed tumor dynamics
Patient-067's MRI follow-up over 167 weeks and 20 scans, with segmented tumor compartments, measured 3-D volumes and expert RANO ratings. The reference view is held constant so interval size change is not confounded by changing scan orientation.
Chemoradiation analysis
Chemoradiation analysis from any observed scan, shown on the patient's own tumor anatomy. Radiation dose and alkylating-agent exposure can be varied across the supported range and followed over 24 weeks.
Surgical planning
Pre-operative analysis on the patient's week-156 anatomy. Extent of resection, depth and cranial access trajectory can be varied; postoperative behavior is referenced to her three earlier observed resections and compared with the fourth postoperative course.
Comparative intervention analysis
Chemoradiation, surgery and combined treatment are compared from the same measured week-156 tumor state for a selected objective and horizon. ABP reports no leading option when the calculated outcomes are not separable.
Matched clinical states, divergent courses
Patients 071 and 072 were closely matched at one scan: similar volume, the same RANO rating, and among the closest 1% of scan pairs on imaging features. Their observed courses diverged: overall survival of 17 weeks after the matched state for one and 96 weeks for the other. Each tumor is shown in that patient's own anatomy at every scan.
Response trajectories
Patients whose tumors responded on imaging (RANO partial or complete response): when each response began, how the compartments changed, how long it lasted, whether the scans before it gave any sign, and a matched patient who progressed instead.
Evidence from the type 1 diabetes study.
In 12 people with type 1 diabetes (194,010 five-minute observations), a model trained on the pooled population and models built from each person's own history were given the same task: find an intervention that moves high blood glucose toward a target.
of tests gave the identical intervention
The population-trained model and the person's own model chose the same intervention in only 24.5% of evaluations. In the other 75.5%, they disagreed about what would move that person toward target. This measures agreement, not which answer was right.
of 400 high-glucose cases moved toward target under ABP inversion
Queried in reverse under clinical constraints, ABP found an intervention direction that reduced the target gap in 96.0% of 400 high-glucose evaluations, with an average gap reduction of 87.8%.