How AI is making an impact at SickKids
AIM supports research and clinical projects that can maximize clinical impact to improve health outcomes for patients and families at SickKids and beyond. These are just some of the ongoing AI projects being led by our SickKids innovators.
Predicting cardiopulmonary arrest in critically ill children
Cardiopulmonary arrest (CPA) is a devastating complication in critically ill children associated with death and significant disability amongst survivors. The only effective treatment is cardiopulmonary resuscitation (CPR). The most effective way to improve outcomes for this population of critically ill children would be to try to prevent CPA altogether. Effective prevention requires an ability to reliably predict the likelihood of the event occurring. SickKids researchers are applying machine learning to this problem using convolutional and recurrent neural networks and continuous physiological signals captured from the patient bedside. Predicting this rare event is a challenging problem that occurs in a complex medical environment full of confounders. The ensemble risk model predicting this event is being deployed in a silent trial to evaluate model performance before full-scale deployment as a clinical risk mitigation tool.
Detecting cancer with whole body MRIs
Early detection of cancer is critical to improving the odds of survival. The team is evaluating how whole-body MRIs (called wbMRI) can be used to detect early-onset cancers in children. SickKids is working with the Children’s Hospital of Philadelphia (CHOP) to use machine learning to improve cancer screening programs in early childhood. The imaging technology is currently only used in adults. Unique challenges of paediatric patients – like changing bone density and difficulty scanning young children – make it hard to obtain clear images. The team is testing how machine learning can help clinicians detect and discover cancer early to reduce the impact of aggressive treatments. This work is done with the generous help of the Mark Foundation for Cancer Research.
Engaging children and youth in AI in paediatric health care
Understanding the perspectives of young people and developing supports to meet their needs is crucial to an ethical, sustainable integration of AI into paediatric healthcare. A SickKids team is reaching out to young people with different experiences to understand what they know now, what they hope for and value, and what they’re concerned about in order to start thinking about how we can better deliver AI-enabled care for paediatric patients. The goal of the project is to use this knowledge to create materials for education, enable consent and assent, and engage patients and families more meaningful in our work at SickKids.
Machine Learning Operations
There are many exciting AI projects across the hospital and Research Institute at SickKids, with tremendous potential to impact paediatric medicine. Currently, research and clinical teams individually design, procure, build and operate infrastructure per AI project/solution. To make this effort scalable, AIM is leading the development of a consolidated Machine Learning Operations (MLOPs) environment that will allow rapid model development, iteration and then deployment for evaluation. The team that designs, builds and maintains the environment includes subject matter experts in technology (data engineers, systems engineers, security engineers and network engineers), ethics, privacy, legal and procurement, as well as other domain experts. To support the MLOPs environment, AIM is also developing a governance model that will balance risk mitigation and efficiency with an environment that promotes rapid innovation and adoption.
Providing faster care in emergency medicine with AI
Wait times in emergency departments (EDs) have increased over 17% in the past 5 years in Canada, contributing to delays in diagnosis and treatment administration. Innovation in health-care delivery is needed to support EDs that face the challenge of providing quality care to increasing volumes of patients with relatively limited provider resources. One common strategy for increasing patient flow, especially in adult EDs, is the use of nursing medical directives at triage. These directives enable nurses to order specific investigations such as bloodwork, urine testing and chest radiographs (CXR) before the patient is seen by a physician.A SickKids team is working to implement machine learning-based medical directives that can identify patients with common diagnoses and issue medical directives in advance of a physician assessment. The goal of the project is to streamline clinical care pathways and automate some processes so clinicians can focus on providing care to their patients.
Thyroid cancer classification tool to better pre-surgical diagnosis
Papillary thyroid carcinoma (PTC) is the most common endocrine malignancy. PTC is also the most common malignancy overall in young women 15-29 years of age. The annual incidence is rising worldwide, with a female predominance emerging in early adolescence. Although thyroid nodules in children are significantly more likely to be malignant than in adults, even in children, most thyroid nodules are benign (about 70%), clinicians are challenged with stratifying patients most likely to harbour malignancy and to prioritize them for biopsy and surgery, while limiting exposure to the often life-long surgical risks among those with benign nodules. Currently, no absolute predictors of benign disease exist. As such, according to historical practice, the accuracy of determining who should undergo a biopsy is only 65.67% (with a false positive rate of 82.14%), and the accuracy of determining who should undergo surgery is only 45.28% (with a false positive rate of 69.05%). A SickKids team is using Random Forests (RF), a type of machine learning method that offers interpretability while retaining predictive power of other ML algorithms, to better classify thyroid cancer. While the advantages of applying RF in the medical realm are clear, the team pushes RF application a step further. Their work is the first to approach this task utilizing demographic, ultrasound and biopsy data already collected from patient records. Their computational pipeline selects only the most statistically important variables for the construction of the final model, a model that transforms the random forest into an even more interpretable rule. The project aims to broaden clinical adoption of such models to aid in the determination of the need for biopsy and surgery, reduce the burden of unnecessary thyroidectomy, and improve quality of life for the population of patients with thyroid nodules.
Machine learning-based tools to assess Ulcerative Colitis severity
Ulcerative colitis (UC) is the most prevalent form of inflammatory bowel disease (IBD) and is associated with decreased quality of life and extensive morbidity. The current diagnosis and treatment of UC is costly and labour intensive, with a variety of methods currently used to assess disease severity. Image-based machine learning techniques can help to automate the scoring of UC severity by quantifying the distribution of inflammatory cells and assigning research-standard scoring indices. SickKids’ AI team (AIM) has developed multiple tools that will help pathologists and gastroenterologists better diagnose and treat their patients with UC.
Predicting obstructive hydronephrosis
Hydronephrosis is a common prenatal ultrasound finding (1-5% of fetuses) showing dilatation of the kidney. Outcomes for these patients vary widely. The majority (70%) of these patients resolve without intervention while the remainder require medical and/or surgical intervention to avoid long-term renal damage. Currently, to determine which patients to intervene on, ultrasound and invasive tests are performed and reviewed by specialists in repeated clinic visits, making hydronephrosis a common and costly condition to manage. To assist in this decision-making, a SickKids team has created a tool to automatically access ultrasounds from these patients to determine if surgical intervention will be indicated. This tool is being tested for prospective accuracy at SickKids as well as for its use in other institutes, including Children’s Hospital of Philadelphia (CHOP) and Stanford Children’s Hospital.