Search

The integration of Artificial Intelligence (AI) in healthcare signifies a substantial shift, offering benefits to patients and healthcare systems while also introducing new risks. The emphasis on patient safety and performance standards is pivotal, especially with the European Union's strides towards regulating AI through the AI Act. This act focuses on classifying AI systems based on risk levels, mandating stringent requirements for high-risk AI, enhancing transparency, and ensuring ethics in AI applications. The concept of an "AI passport" is introduced as a living document detailing the AI system's purpose, ethical declarations, training, evaluation, and potential biases. This passport aims to enhance transparency and safety in medical AI applications, serving as a comprehensive record for patients, clinicians, and stakeholders. The AI passport, structured in JSON format, encapsulates key information about the AI system as a mechanism for continuous performance evaluation and transparency. This initiative may represent a significant step towards mitigating the risks associated with AI in healthcare, emphasizing the importance of accountability, transparency, and patient safety in the development and application of AI technologies.

Artificial intelligence (AI)-based clinical decision support systems are gaining momentum by relying on a greater volume and variety of secondary use data. However, the uncertainty, variability, and biases in real-world data environments still pose significant challenges to the development of health AI, its routine clinical use, and its regulatory frameworks. Health AI should be resilient against real-world environments throughout its lifecycle, including the training and prediction phases and maintenance during production, and health AI regulations should evolve accordingly. Data quality issues, variability over time or across sites, information uncertainty, human-computer interaction, and fundamental rights assurance are among the most relevant challenges. If health AI is not designed resiliently with regard to these real-world data effects, potentially biased data-driven medical decisions can risk the safety and fundamental rights of millions of people. In this viewpoint, we review the challenges, requirements, and methods for resilient AI in health and provide a research framework to improve the trustworthiness of next-generation AI-based clinical decision support., (©Carlos Sáez, Pablo Ferri, Juan M García-Gómez. Originally published in the Journal of Medical Internet Research (https://www.jmir.org), 28.06.2024.)

Most learning algorithms for classification use objective functions based on regularized and/or continuous versions of the 0-1 loss function. Moreover, the performance of the classification models is usually measured by means of the empirical error or misclassification rate. Nevertheless, neither those loss functions nor the empirical error is adequate for learning from imbalanced data. In these problems, the empirical error is uninformative about the performance of the classifier and the loss functions usually produce models that are shifted to the majority class. This study defines the loss function L BER whose associated empirical risk is equal to the BER. Our results show that classifiers based on our L BER loss function are optimal in terms of the BER evaluation metric. Furthermore, the boundaries of the classifiers were invariant to the imbalance ratio of the training dataset. The L BER-based models outperformed the 0-1-based models and other algorithms for imbalanced data in terms of BER, regardless of the prevalence of the positive class. Finally, we demonstrate the equivalence of the loss function to the method of inverted prior probabilities, and we define the family of loss functions L WER that is associated with any WER evaluation metric by the generalization of L BER.

In the last decade, machine learning (ML) techniques have been used for developing classifiers for automatic brain tumour diagnosis. However, the development of these ML models rely on a unique training set and learning stops once this set has been processed. Training these classifiers requires a representative amount of data, but the gathering, preprocess, and validation of samples is expensive and time-consuming. Therefore, for a classical, non-incremental approach to ML, it is necessary to wait long enough to collect all the required data. In contrast, an incremental learning approach may allow us to build an initial classifier with a smaller number of samples and update it incrementally when new data are collected. In this study, an incremental learning algorithm for Gaussian Discriminant Analysis (iGDA) based on the Graybill and Deal weighted combination of estimators is introduced. Each time a new set of data becomes available, a new estimation is carried out and a combination with a previous estimation is performed. iGDA does not require access to the previously used data and is able to include new classes that were not in the original analysis, thus allowing the customization of the models to the distribution of data at a particular clinical center. An evaluation using five benchmark databases has been used to evaluate the behaviour of the iGDA algorithm in terms of stability-plasticity, class inclusion and order effect. Finally, the iGDA algorithm has been applied to automatic brain tumour classification with magnetic resonance spectroscopy, and compared with two state-of-the-art incremental algorithms. The empirical results obtained show the ability of the algorithm to learn in an incremental fashion, improving the performance of the models when new information is available, and converging in the course of time. Furthermore, the algorithm shows a negligible instance and concept order effect, avoiding the bias that such effects could introduce., (Copyright © 2011 Elsevier Inc. All rights reserved.)

Justification: Automatic brain tumor classification by MRS has been under development for more than a decade. Nonetheless, to our knowledge, there are no published evaluations of predictive models with unseen cases that are subsequently acquired in different centers. The multicenter eTUMOUR project (2004-2009), which builds upon previous expertise from the INTERPRET project (2000-2002) has allowed such an evaluation to take place., Materials and Methods: A total of 253 pairwise classifiers for glioblastoma, meningioma, metastasis, and low-grade glial diagnosis were inferred based on 211 SV short TE INTERPRET MR spectra obtained at 1.5 T (PRESS or STEAM, 20-32 ms) and automatically pre-processed. Afterwards, the classifiers were tested with 97 spectra, which were subsequently compiled during eTUMOUR., Results: In our results based on subsequently acquired spectra, accuracies of around 90% were achieved for most of the pairwise discrimination problems. The exception was for the glioblastoma versus metastasis discrimination, which was below 78%. A more clear definition of metastases may be obtained by other approaches, such as MRSI + MRI., Conclusions: The prediction of the tumor type of in-vivo MRS is possible using classifiers developed from previously acquired data, in different hospitals with different instrumentation under the same acquisition protocols. This methodology may find application for assisting in the diagnosis of new brain tumor cases and for the quality control of multicenter MRS databases.

(1)H MRS is becoming an accurate, non-invasive technique for initial examination of brain masses. We investigated if the combination of single-voxel (1)H MRS at 1.5 T at two different (TEs), short TE (PRESS or STEAM, 20-32 ms) and long TE (PRESS, 135-136 ms), improves the classification of brain tumors over using only one echo TE. A clinically validated dataset of 50 low-grade meningiomas, 105 aggressive tumors (glioblastoma and metastasis), and 30 low-grade glial tumors (astrocytomas grade II, oligodendrogliomas and oligoastrocytomas) was used to fit predictive models based on the combination of features from short-TEs and long-TE spectra. A new approach that combines the two consecutively was used to produce a single data vector from which relevant features of the two TE spectra could be extracted by means of three algorithms: stepwise, reliefF, and principal components analysis. Least squares support vector machines and linear discriminant analysis were applied to fit the pairwise and multiclass classifiers, respectively. Significant differences in performance were found when short-TE, long-TE or both spectra combined were used as input. In our dataset, to discriminate meningiomas, the combination of the two TE acquisitions produced optimal performance. To discriminate aggressive tumors from low-grade glial tumours, the use of short-TE acquisition alone was preferable. The classifier development strategy used here lends itself to automated learning and test performance processes, which may be of use for future web-based multicentric classifier development studies., (Copyright (c) 2008 John Wiley & Sons, Ltd.)