OVERVIEW

With the rise of technology and computing power in smartphones, personal desktop computers, and cloud-based services, complex Algorithms, Artificial Intelligence (AI) and Machine Learning (ML) have moved rapidly from the realm of theory into actual products with widespread use. Flowbit understands our clients’ interest in the application of these technologies in medical science and engineering. We bring together ideas, techniques, and processes from disparate research papers, journal articles, and books and offer a multi-discipline-based systematic engineering approach.

Flowbit understands that to solve complex healthcare problems, we require many algorithms to work in concert together. As such, we design our systems so that solving complex problems does not result in complex dependencies and algorithms. We do so by ensuring our design is modular and broken down into smaller easy-to-understand processing steps. We also provide you with visibility into how the overall algorithm is performing, and how it transforms your data— thereby avoiding a black box approach to developing AI. To demonstrate this approach, we have provided our step-by-step approach to performing analysis on Audio and Image processing below.

Using various numerical methods and software libraries, we can help you analyze signal data by performing conversions from raw sensor formats, followed by quality control steps that smooth out noise, determine and apply thresholds, and detect edges and regions of interest. These steps usually occur before the data is fed into another algorithm for processing.

"CL Society 409: Child in a ceremony" by francisco_osorio is licensed with CC BY 2.0.

Using machine learning we can provide accurate diagnosis and predict health conditions. Flowbit will help you design the right model architecture for your use case and data— employing various modeling techniques including Perceptrons, Convolution, Recurrent, or Long/Short Term Memory Neural Networks.

In the following example, we have two sets of audio data of heart sounds collected from a digital stethoscope, i.e., normal and abnormal heart sounds. An AI model can be trained via supervised learning to infer normal or abnormal heart sounds using this dataset. Essentially, the audio data is converted from the temporal domain to a frequency domain, so that a normalized frequency power analysis algorithm can be applied. Using these pre-processed data frame sets, an AI model can be trained to distinguish normal from abnormal heart sounds even when the differences are subtle.

When it comes to Verified and Validated Algorithms for AI/ML, as well as preprocessing steps, we run the same code on iOS and Android to give you consistent results on the two different platforms. We aim to reduce any variation in results between the Cloud and mobile platforms you use, by utilizing Tensorflow converters and their openly available conversion techniques.

Underneath the hood, we utilize C/C++ programming language exclusively for executing algorithms that perform pre-processing of data and inferring of results. The C/C++ code base we design and develop is reusable and cross-compiled for iOS, Mac OS, MS Windows and Unix-based operating systems. This eliminates programming differences. In addition, C/C++ almost always gives us faster execution times, as there is less overhead and fewer intermediate layers of instruction processing between the code and hardware it runs on.

While U.S. FDA SaMD device reviews were not traditionally designed for adaptive Algorithms for Artificial Intelligence/Machine Learning (AI/ML), the FDA provides guidance including April 2019 discussion paper, “Proposed Regulatory Framework for Modifications to Artificial Intelligence/Machine Learning-Based Software as a Medical Device” As well as January 2021, the FDA released an AI/ML Action Plan based on this discussion paper.

We have deep expertise in planning and executing verification and validation testing that meets the FDA’s traditional regulatory framework for SaMD. In addition, we work within the new Proposed Regulatory Frameworks by FDA that are under consideration for adoption, such as Algorithm Change Protocol (ACP) in the context of the Total Product Lifecycle (TPL). These and other measures help the FDA to establish reasonable trust of assurance on safety and effectiveness of the medical device, while embracing the iterative improvement power of AI and ML.

A robust AI/ML-engineered enterprise solution must contain a host of activities that include Machine Learning Development practices, in addition to Modern Software Development practices that still apply.

Thanks to some great AI/ML platforms, such as TensorFlow, CoreML, MATLAB, and PyTorch the actual machine learning code development is only but a small fraction of the effort involved in developing the AI/ML system.