Data analysis app for medical garment usage prediction

Type A set of scripts
TECHNOLOGIES Google BigQuery, Python, TensorFlow
AREAS OF EXPERTISE Data Science, neural networks, Google services integration, Big data
TEAM 1 developer, 1 project manager

We were contacted by a client who had gathered a lot of data on the storage and movement of dispensers with garments for medical staff. The client wanted to use this data for the optimization of garment placement and movement processes for hospitals in Europe, which we successfully achieved as the result of this project.

All hospital employees should wear specific work clothes. These clothes are stored in dispensers. Every employee uses multiple types of cloth: gowns, gloves, T-shirts, etc. After the work is done, the clothes are put in a container. Since every piece of work clothes has an RFID tag, its usage can be logged by the tracking system: who took it, when, where, and when it was released.

The problem was – a lack of some types of cloth and excess of another. As a result, employees were unable to acquire all the necessary garments set in one dispenser and had to move to another part of a hospital to get it.

  1. First of all, we have performed a deep data analysis of the provided data. We have found this data to be very noisy with many bulk transactions, lost items, etc. It seemed inappropriate for further investigation.
  2. Predict garments usage by medical staff: when, where and in what amounts it would be required for the upcoming shifts.
  1. We have studied the context of the garments rotation process and realized that employees release their clothes using container nearest to the last workplace, no matter where theses clothes were picked up initially. This simple idea became a key to understanding the data. And so we tracked all cloth returns back in time to find corresponding withdraws. Then we filtered the obtained data by dropping items with usage time for more than two weeks. Thus we obtained a full picture of the real demand for every type and size of cloth on every container.
  2. We built a model using CNNs (convolutional neural networks) on Tensorflow, which receives a two-weeks usage history and predicts usage of every type of garment for every day of the pending week. Internally, it converts the log into a [14xN] matrix (N is a number of garment models) and passes it to CNN. This CNN also contains down- and up-sampling layers. It produces a 7xN matrix of predicted demand for every day and a garment model.




Our model successfully predicts the usage of inventory work clothes by employees of an enterprise for the next week using two-weeks exploitation history as an input (which is around 2 million records) and supports the business processes of medical organizations in Europe.

AR Office application
AR Office application
ML optimization of market conversion for a financial aggregator
ML optimization of market conversion for a financial aggregator