AI

Recent medical large language models have been reported to approach physician-level performance in selected diagnostic and prescribing-related tasks, yet regulations continue to restrict clinical authority to licensed human professionals. Without assuming artificial general intelligence (AGI), I propose a Clinical AI License with four requirements, disease-domain performance thresholds, staged evaluation analogous to clinical trials, graduated scope of practice, and ongoing monitoring, together with an allocation of primary legal liability to the licensed operating entity. Responsibility, empathy, and safety are not fixed human monopolies; under an appropriate licensing framework, they can be specified, evaluated, and continuously monitored in AI-based clinical systems, potentially enabling scalable and more equitable healthcare.

Accurate remaining useful life (RUL) prediction of aero-engines under varying operating conditions is challenging because condition-induced sensor variations can obscure degradation signatures. This study evaluates a multiscale dual-attention residual network (MS-DA-ResNet) under a leakage-controlled protocol on the multiple-condition FD002 and FD004 subsets. The 125-cycle cap is applied only to training and validation targets, whereas official NASA test endpoint labels remain uncapped. All models use the same engine-level split, fixed 14-sensor input, causal preprocessing, and best-validation checkpoint, as well as three identical matched seeds. Under the corrected official-label evaluation, MS-DA-ResNet provides the best aggregate results on both datasets: FD002 RMSE 25.14 ± 0.41, MAE 16.71 ± 0.21, R² 0.781 ± 0.007, and NASA Score 5597 ± 389; FD004 RMSE 24.58 ± 0.22, MAE 17.35 ± 0.07, R² 0.797 ± 0.004, and NASA Score 3430 ± 77. MS-DA-ResNet remains competitive but does not achieve universal superiority. Its error-direction behavior is interpreted separately, using Bias and the overestimation ratio. Condition-wise normalization with causal smoothing improves FD004 relative to the tested global-normalization and no-smoothing alternatives. A corrected five-seed L = 30 sensitivity run does not improve the primary L = 15 configuration.

AI systems that measure lesions on request can compute correctly on the wrong lesion, an outcome no benchmark I found scores. I located where such mix-ups arise, in three stages on public lung CTs. A stratified audit of 1785 DICOM segmentation series found identical coded semantics for all 5192 pairs of different nodules on one scan; nodule numbers existed but needed linking to the request. A gate trusting structural checks authorized all 200 constructed requests, 89 for a mismatched and 41 for an ambiguous target. An expert wrote 60 requests in three sets, with each design frozen first. A rule parser carried a misread rank (“longest diameter” read as “largest”) in every resolution; removing it post hoc kept 13 of 26 correct resolutions, and agreement with a choice-constrained model parser removed all 25 rank-dependent wrong ones. A whole-lung definition of “lateral” excluded an intended nodule. In the confirmatory set, both pre-specified consensus gates authorized the nearest unintended nodule in two of 20 requests. Adjudication found three requests naming a lobe the images contradicted and two unassignable lobes; no map error was demonstrated. Parser agreement did not catch a request–image discordance. Clinical safety and workflow benefit were not evaluated.