Services: Artificial intelligence
Client: A UK-based company providing services for healthcare organizations
Project: predicting demand for cancer treatment in hospitals
About the Client
The Client is a UK-based company that provides leading services to healthcare organizations in the United Kingdom advising healthcare institutions on improving their operational performance.
The Client saw an opportunity for using AI technologies to predict demand in hospitals for cancer treatment as well as other patient needs. With predictions of demand, hospitals would be able to better prepare their capacity for periods of higher demand, which would result in better quality of service, less strain on healthcare staff, and, most importantly, faster diagnosis allowing to save more lives. The challenges the Client was facing were in finding the fastest and most efficient way of leading patients through the process of Cancer Pathways from referral to diagnosis and predicting treatment demand that would allow healthcare organizations to better plan resources.
Validating the hypothesis that cancer treatment demand can be predicted accurately
The main goal of the project was to validate the initial hypothesis that a model predicting cancer treatment demand and patients’ needs can be built to support healthcare institutions in their operations.
Selecting a model that would best fit the criteria specified by the Client
Building the right machine learning algorithm is a crucial element of AI adoption. It’s essential that various models are considered and tested, and the one that fits the requirements specified by the client is selected.
Delivering custom data strategy with recommendations for AI adoption
AI adoption requires a strategic approach – so custom data strategy is prepared along with recommendations of steps to be taken to provide for more seamless adoption of AI technologies.
Inefficient strategies of planning resources in hospitals
Hospitals struggle to plan and organize their resources in accordance with demand, not only in terms of cancer treatment but generally patient needs. When understaffed or lacking essential resources, medical staff cannot provide the right standard of care for their patients. If given the opportunity to learn about upcoming increase in demand, they would be able to plan better, making their job more manageable and patient care of higher quality.
Guiding patients through Cancer Pathways currently takes too long
Getting each patient through Cancer Pathways from referral to diagnosis is extremely important. The process can consume quite some time when not managed appropriately.
Use of data from other (non-NHS) sources
To provide the best possible predictions, other data (coming from sources different from NHS) proves valuable. Other data that could be used includes TV & radio campaigns, social media, etc.
Conducting an AI Sprint
AI Sprint is the first step we take with all clients wanting to implement artificial intelligence in their companies. The AI Sprint is a combination of a two-day workshop and PoC (proof-of-concept). Thanks to such an approach, our clients have insight into all the opportunities and risks involved in adopting AI and can see the probability of success of AI adoption in their organization. Thorough analysis of their business goals, processes, requirements, and data allows us to provide the client with custom recommendations for further steps to allow for a well-informed decision about AI investments.
Performing exploratory data analysis
We enriched the data obtained from NHS, which included patients’ age, gender, the date of issuing the referral, and suspected tumor group, with other data including information about the weather, bank holidays and other holidays, sports events, cancer-themed posts in social media and keyword searches. All data was then divided into 2 sets: training data and testing data. The training data is used to “teach” the model, and the testing data is later used to check whether the model works on new data.
Validating the hypothesis
The goal of the AI Sprint is to validate the hypothesis that a given objective can be achieved with AI. Thanks to that, the client gains much valuable knowledge about how AI can work in their organization and is provided with the PoC that paves the path for AI development and further adoption.
Python and libraries for data science, including NumPy, SciPy, Scikit-learn, TensorFlow; Google Cloud Platform
Data hypothesis validated within one month
During the AI Sprint, our data science team analyzed the available data and considered various models that could potentially solve the challenge that the client was facing. In just three weeks, we conducted a full AI Sprint that ended with a positive validation of the hypothesis meaning that it is possible to build a model predicting demand for cancer treatment with the use of the collected NHS and publicly available data.
Appropriate model identified
Identifying the right model to achieve the goal is imperative. Every model works differently and will bring different results. In order to select the appropriate model, our team tested various models that had the potential to achieve the set goal. At the end of the AI Sprint, we present the results of the models and point the most successful one. At this point, the model can be further improved to be later successfully implemented.
Recommendations for AI implementation
We provided the client with recommendations for further steps in their AI adoption: suggestions of actions to be taken in order to implement AI successfully. We also assessed the probability of full AI adoption’s success in the Client’s company, which was around 74%. With this knowledge, we were able to list the next significant steps in the process and estimate the value of the product once fully implemented.
One of the models built during the AI Sprint proved to be a solution to the Client’s problem – it was able to generate predictions for cancer treatment demand for the upcoming 28 days with constant precision. The model proved to be able to mimic real behavior. Being provided with more observations, the model’s precision can be further improved.