AI at the bedside: heart failure, imagined speech and CT calibration

Heart failure strains care systems. Patients are older, carry other conditions and often end up in hospital. Remote monitoring promises to catch deterioration earlier. But telemonitoring data are sparse, sampled irregularly and dominated by cases with no event.

Detecting deterioration from telemonitoring data

TRACER, a transformer with contrastive event representation, was built for exactly this kind of data. It combines time-aware biomarker embeddings, contrastive pretraining that sharpens anomaly detection and independent binary classifiers for each finding. The authors worked on remote measurement sequences from 276 heart failure patients. They split the sequences into overlapping windows along time rules and labelled each window by whether hospitalisation fell at its end. The model correctly predicted 66.7 percent of the pathways leading to hospitalisation for heart failure, with an over-prediction of 7.9 percent, in a strongly imbalanced real-world set. Reframing training as an event detection task also improved results over training directly on prognostic windows and made better use of the few events available. The authors stress that the aim is to generate alerts that allow early intervention, not to replace a clinical decision.

Imagined speech and honest evaluation

Decoding imagined speech from electroencephalography is being explored as a communication route for people with severe motor impairment. The trouble is that reported results often rest on evaluation protocols that do not reflect generalisation to new people. A Danish team led by Frederik Møllskov Trier ran a transparent baseline study on a multiclass EEG imagined speech dataset under a strictly subject-independent evaluation scheme.

Two processing and feature extraction pipelines were compared: statistical features in the time domain and power in spectral bands. The second path reached a significantly higher mean accuracy at trial level, 49.03 plus or minus 4.18 percent against 37.97 plus or minus 3.79 percent, in coarse between-subject classification. Forward feature selection showed that most of the discriminative information sits in a limited subset of bands. The authors treat this as a solid basis for further work on brain-computer interfaces, not as a finished product.

Calibration when the model is a black box

The third paper deals with the safety of radiology systems. Decisions about triage, follow-up imaging or treatment planning are only safe if the probabilities a model returns are calibrated. They have to reflect real risk. Standard techniques for improving calibration, such as Monte Carlo dropout or deep ensembles, need access to model parameters or retraining. Commercial clinical systems are black boxes.

The proposed framework is model-agnostic. It uses clinically justified test-time augmentation, applying geometric and physics-inspired perturbations of three-dimensional computed tomography, and learns a strategy for aggregating probabilities without access to the inside of the model or to training data. On pulmonary embolism and intracerebral haemorrhage detection tasks, the DualTTA variant achieved the best calibration among test-time augmentation methods. It cut the expected calibration error by 54 percent, from 0.239 to 0.109, and by 43 percent, from 0.051 to 0.029. In most metrics it also beat techniques that require access to the inside of the model, such as temperature scaling, Monte Carlo dropout and deep ensembles.

Together the three papers mark a practical boundary: the place where a model works on real, sparse and irregular data, and the place where the added value comes not from a new model but from honest evaluation and calibration.