An integration engineering framework for machine learning in healthcare

Assadi, Azadeh; Laussen, Peter C.; Goodwin, Andrew; Goodfellow, Sebastian D.; Dixon, William G; Greer, Robert O.; Jegatheeswaran, Anusha; Singh, Devin; McCradden, Melissa D; Gallant, Sara N.; Goldenberg, Anna; Eytan, Danny; Mazwi, Mjaye

doi:10.3389/fdgth.2022.932411

Cited by 9 publications

(2 citation statements)

References 34 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Results: Data from 9415 patients who underwent cardiac surgery (median [IQR] age, 68 [60][61][62][63][64][65][66][67][68][69][70][71][72][73][74] years; 2554 [27.1%] women) were included. Overall mortality rates at 30 days, 1 year, and 5 years were 268 patients (2.8%), 420 patients (4.5%), and 612 patients (6.5%), respectively.…”

Section: Main Outcomes and Measuresmentioning

confidence: 99%

“…Integration into an already existing clinical workflow is, and will remain, the challenge faced by researchers, clinicians, and others in the implementation of decision support systems 71,72 It has been suggested that the American National Aeronautics and Space Administration (NASA) system of Technology Readiness Levels (TRLs), which assesses air-and spacecrafts "readiness to fly", could also be adopted for digital health innovations. Following this principle, clinical applications of AI-enabled decision support systems should occur only when there is adequate evidence of its safety and efficacy 70,73 .…”

Section: Supplementary Materialsmentioning

confidence: 99%

See 1 more Smart Citation

Applications of Machine Learning in Anaesthesiology and Critical Care

Alves Castela Cardoso Forte

View full text Add to dashboard Cite

Since ML techniques are based on "conventional" statistics, it is important to dive deeper into the process of training 2 . The collected data is usually split into two parts -the so-called training and testing sets -allowing for a successfully trained algorithm to make predictions on data it has not seen before. This is a crucial step, as ML algorithms instantly identify the particularities of a specific dataset, which may largely shape the resulting prediction model. This phenomenon, called overfitting, results in inaccurate predictions of models when using new data, therefore hampering the model's generalizability. Therefore, a model's performance is assessed based on the "new" predictions made on the "clean" test data that the algorithm has not seen during training, as this provides an estimate of its potential applicability to new datasets. value of a computationally-expensive technique like dynamic time warping (DTW).The third part of this thesis expands beyond mortality prediction and clustering, focusing on applying computer vision to develop a proof-of-concept algorithm to identify sepsis and acute illness. Through one of the first editorials ever published in The Lancet Digital Health, in chapter 8 we reflect on lessons learnt from a landmark prediction study in the ICU, and highlight the essential elements of research for the prediction long-term outcomes in critical care. In particular, gathering of correct data will require changes to data collection strategies and infrastructure, including closer collaboration between clinicians, researchers, data scientists, and national medical data registries. To explore the potential added value of alternative, new data sources in settings like the ICU or emergency room, in chapter 9 we trained a deep learning algorithm on a dataset of simulated and augmented facial photographs of healthy volunteers mimicking the "gestalt" (facial cues) of acutely ill patients. This algorithm was then validated in an external dataset of photographs of volunteers injected with lipopolysaccharide (LPS) to induce a controlled state of "acute illness".Finally, in chapter 10, the main findings of this thesis are outlined and discussed.

show abstract