At first glance, a blood test result is a table of numbers: hemogram values, biochemistry parameters, reference ranges. Yet inside that table lives something you cannot see by reading values one at a time: the patterns between parameters. An experienced physician reads the picture that several values paint together, almost intuitively. The promise of artificial neural networks in this space is exactly that: capturing those patterns systematically. In this article we look at why feedforward neural networks — the FNN architecture — are such a natural fit for blood test data.
What is an FNN and how does it work?
The feedforward neural network is the most fundamental architecture in neural computing. Data flows in one direction, from the input layer to the output layer; the artificial neurons in each layer weight and transform the values arriving from the previous layer and pass them on. During training, those weights are updated step by step so that the gap between the model's predictions and the actual outcomes shrinks. The architecture's strength is its simplicity: when the input is structured numerical data like a blood panel — rather than complex inputs like images or free text — an FNN with properly scaled inputs can produce remarkably effective results.
Why blood test data suits neural networks
From a machine learning perspective, a blood panel is close to an ideal data format. Consider why:
- Fixed, well-defined inputs: every parameter has a known name, unit and reference range; the model sees identically structured data in every sample.
- Numerical nature: the values are already numbers; no text parsing or image processing layer is needed.
- Richness of patterns: clinical meaning often hides not in a single parameter but in how parameters move together. Capturing these multidimensional relationships is the natural strength of neural networks.
There is, of course, another side to the coin: the patterns a model learns are bounded by the quality and representativeness of its training data. Measurement differences between devices, reference ranges that vary across populations, and missing values are engineering problems that demand careful work.
Our project in development: FNN Blood Analysis
At Afinitem we are running an R&D project that puts this approach into practice. FNN Blood Analysis aims at automatic analysis of blood test results using feedforward neural networks and deep learning models. Let us underline the status: the project is in development; it is not a released diagnostic tool. Our goal is not a system that decides in the physician's place, but a support layer that highlights noteworthy patterns in test results and contributes to the clinician's assessment. We believe this is what responsible medical AI looks like: capabilities validated step by step, rather than ambitious marketing.
Training: from data preparation to validation
An FNN's success is decided in data preparation before architecture. Bringing values from different laboratories onto a common scale, handling missing parameters consistently, and keeping outliers from misleading the model are homework that precedes training. During training, the critical discipline is testing on data the model has never seen: a model that shines on its training set but falters on held-out validation data has memorized, not learned. Explainability is just as indispensable — flagging a result as noteworthy is not enough; the model must show which parameters drove that assessment, because a physician cannot and should not trust a signal whose rationale is invisible.
The meeting of neural networks and laboratory data is on course to become a natural part of routine workflows in the coming years. To follow our work in this area or send us your questions, you can use our contact page.