Modules
Enhanced, Personalized and Integrated Care for Infection Management at Point of Care (EPiC IMPOC) is a CE marked decision support system designed to record and integrate clinical, physiological, and imaging data at the patient’s bedside on hand-held computing devices while providing instant decision-making assistance to health care workers at the point of care.
EPiC IMPOC's modular design breaks down the system to reduce its complexity by encapsulating functionality and processes within reusable modules. An overview of these modules, ranging from patient administration and data collection to prognostic functionality is presented below.
Case-Based Reasoning
Reuse and transfer clinician's knowledge to make better-informed decisions!
In clinical environments, physician reasoning is based on knowledge acquired from past cases personally experienced which is exactly what CBR does!
​
The aim of CBR is to solve new problems based on the solutions of similar past problems in form of cases. CBR is considered a methodology to follow rather than an algorithm in itself. Initially, the CBR methodology requires a set of cases or training examples denoted as case-base. Then, for a specific query, the CBR cycle is divided into four different phases.
​
-
Retrieve: Given a query patient, this phase retrieves from the database those cases that are relevant, based on any predefined similarity metric.
-
Reuse: propose a suitable solution for the query patient by adapting and/or combining the solution of the retrieved cases.
-
Revise: experts monitor the evolution of the patient to assess whether or not the adapted/proposed solution was appropriate.
-
Retain: determine whether or not the patient is of utility to be added into the case-base for future queries.
Communication
Promotes communication between nurses, clinicians and infection specialists
Data collection
Improves homogeneous data collection of clinical symptoms and vital signs
Knowledge
Facilitates comparison with previous similar patients and their outcomes
Decision Making
Supports clinical decision making and encourages patient monitoring
Heterogeneity often appears in healthcare data because patients' populations, environment, procedures, and treatment protocols in medical centres vary. In addition, this data is usually of high complexity and its revision in standard format (e.g forms) is very time consuming. Thus intelligent and intuitive visualisation approaches and performance optimisation for patient similarity retrieval are key!
In this section we introduce the similarity retrieval module which allows to identify clinically meaningful projections of such high dimensional data in two dimensions (often called latent space) that (i) align with the established clinical characteristics of a particular disease and its progression and (ii) maps patients with similar manifestations/phenotypes close together in the latent space.
Autoencoders are a type of neural network which aim to learn an encoding of the input data. They do so by attempting to copy their input to their output, having gone through a hidden layer h which has fewer neurons than the input has features. This hidden layer, often called a bottleneck, forces the model to extract the essential features present in the input data to then be able to reconstruct the input as faithfully as possible. An autoencoder with two neurons in its bottleneck can then be used to visualise data in two dimensions!
​
The latent space is simply a representation of compressed data in which similar data points, in our case patients, are closer together in space. The following aspects should be considered when constructing and validating the model:
​
-
Similarity retrieval: To effectively use the model for this purpose, we need to ensure that patients with similar symptoms and phenotypes are represented close together in the latent space.
-
Patient trajectory: To effectively use the model for this purpose, we need to ensure that distances are preserved in the latent space so that it can be used to visualize patient trajectories; that is, their evolution over time.
-
Distance preservation: There are applications where maintaining the ordering of distances is more important than having a linear relationship between the distances in the original and reduced spaces. Similarity retrieval, for example, will provide the same results in the original space and the two-dimensional space if the Spearman rank correlation coefficient of the distances in both spaces is one.
​​​
-
Alignment with disease characteristics: To validate the use of the latent space in clinical practice, it is essential to ensure that it aligns with pathological characteristics of the disease progression so that it can be used in clinical practice. Thus, it is necessary to properly understand how the features, phenotypes, and categories behave in this reduced 2D space.
-
Timothy M. Rawson, Hernandez, Bernard, Luke S. P. Moore, Pau Herrero, Esmita Charani, Damien Ming, Richard C. Wilson, Oliver Blandy, Shiranee Sriskandan, Mark Gilchrist, Christofer Toumazou, Pantelis Georgiou, and Alison H. Holmes. A Real-world Evaluation of a Case-based Reasoning Algorithm to Support Antimicrobial Prescribing Decisions in Acute Care. Clinical Infectious Diseases, 72(12):2103–2111, 2021 [Link]
-
Hernandez, Bernard. Data-driven web-based intelligent decision support system for infection management at point of care. PhD thesis, Imperial College London, Department of Electrical and Electronic Engineering, London, UK, 2018 [Link]
-
Hernandez, Bernard, Pau Herrero, Timothy M. Rawson, Luke S. P. Moore, Esmita Charani, Alison H. Holmes, and Pantelis Georgiou. Data-driven web-based intelligent decision support system for infection management at point-of-care: Casebased reasoning benefits and limitations. pages 119–127, 2017 [Link]
Probabilistic Inference
Tailored and specific advice by providing answers to clinician's most predominant questions using probabilities
In clinical environments, the acute infection management pathway followed by clinicians has been defined as a step-wise Bayesian model in which each step adds systematically information to optimise diagnosis and management of infection.
​
First, they evaluate the physiological parameters of the patient and try to localise and confirm the infection. Then, physicians review and plan further investigations such as pathology laboratory tests, specimen collection for susceptibility testing or imaging. These previous steps allow physicians to construct a clinical picture of the severity of the infection. From this clinical picture, a decision to whether or not to initiate antimicrobial therapy is made with local microbiology guidance as a determinant factor. Finally, an internal and/or external review of the patient is conducted to refine the antimicrobial therapy and the entire process is repeated.​
Assessment
Evaluate the patient to asses what is the likelihood of having an infection?
Location
To plan further investigations, what is the most likely site of infection?
Severity
Will the patient suffer complications in the following days?
Treatment
Should we initiate antimicrobial therapy? which one?
Browse through all the models previously trained, analyse their performance outcomes such as the area under the ROC curve, the area under the PR curve or the probability calibration curve or investigate the SHAP values to understand which features are influencing the predictions. These models are all ready to be used and can be queried through the implemented API so you can evaluate the performance with your own datasets or even use this predictions for clinical decision support.​
-
Timothy M. Rawson, Hernandez, Bernard, Richard C. Wilson, Damien Ming, Pau Herrero, Nisha Ranganathan, Keira Skolimowska, Mark Gilchrist, Giovanni Satta, Pantelis Georgiou, and Alison H. Holmes. Supervised machine learning to support the diagnosis of bacterial infection in the context of COVID-19. JAC-Antimicrobial Resistance, 3(1), 02 2021 [Link]
-
Hernandez, Bernard. Data-driven web-based intelligent decision support system for infection management at point of care. PhD thesis, Imperial College London, Department of Electrical and Electronic Engineering, London, UK, 2018 [Link]
-
Hernandez, Bernard, Pau Herrero, Timothy M. Rawson, Luke S. P. Moore, Benjamin Evans, Christofer Toumazou, Alison H. Holmes, and Pantelis Georgiou. Supervised learning for infection risk inference using pathology data. BMC medical informatics and decision making, 17(1):1–12, 2017 [Link]​
-
Timothy M. Rawson, Esmita Charani, Luke S. P. Moore, Hernandez, Bernard, Enrique Castro-Sánchez, Pau Herrero, Pantelis Georgiou, and Alison H. Holmes. Mapping the decision pathways of acute infection management in secondary care among uk medical physicians: a qualitative study. BMC medicine, 14(1):1–10, 2016 [Link]
AMR Surveillance
Make the most of your susceptibility testing data to understand emerging local antimicrobial resistance patterns
With the increasing electronic recording of data, there is a growing interest in the potential secondary use of microbiology records to provide the necessary information to support antimicrobial stewardship programs. These programs are crucial to guide health care organisations in designing evidence-based policies to combat AMR. In particular, susceptibility reporting has shown to be a determinant data source to inform empiric antimicrobial therapy selection.
Rates
Proportion of resistant isolates for a given set of susceptibility tests
Trends
Evolution of resistance rates over the last years and future perspectives
Multiple AMR
Proportion of antimicrobials to which a pathogen is resistant
Spectrum
Range of microorganisms that can be treated with an antimicrobial
The benefits of providing local rather than general antimicrobial surveillance data of a higher quality is two fold. Firstly, it has the potential to stimulate engagement among physicians to strengthen their knowledge and awareness on antimicrobial resistance which might encourage prescribers to change their prescription habits more willingly. Moreover, it provides fundamental knowledge to the wide range of stakeholders to revise and potentially tailor existing guidelines to the specific needs of each hospital. An AMR report summary report for Escherichia Coli in urine samples is shown below.
-
Hernandez, Bernard, Pau Herrero-Viñas, Timothy M. Rawson, Luke S. P. Moore, Alison H. Holmes, and Pantelis Georgiou. Resistance trend estimation using regression analysis to enhance antimicrobial surveillance: A multi-centre study in london 2009–2016. Antibiotics, 10(10), 2021 [Link]
-
Hernandez, Bernard. Data-driven web-based intelligent decision support system for infection management at point of care. PhD thesis, Imperial College London, Department of Electrical and Electronic Engineering, London, UK, 2018 [Link]