Timely discharge of hospitalized patients can prevent complications and reduce costs. In this paper, we have investigated whether outpatient services, i.e., diagnostic exams or specialist consultations provided by external wards, have a clear impact on Length of Stay ( ). In particular, we have worked on an event log of more than 7000 real patient traces, logging the sequence of outpatient services provided during hospitalization, and we have classified the traces into long ( ≥ 20 days) versus short (< 20 days) adopts a Long Short-Term Memory (LSTM) network. The very high quality of the classification results suggests that outpatient services play a significant role in determining organization.
Length of Stay (
) is a clinical metric that measures the time elapsed between the hospital admission
of a patient and her/his discharge. Unnecessary days in the hospital can lead to increased
hospitalacquired patient complications (e.g., infections) and increased costs for patients and healthcare systems.
Delays in hospital discharge may be related to unnecessary waiting, poor organization of care, delays
in decision-making, or dificulties related to discharge planning [
In Italy, a patient is typically admitted to the ward where s/he will undergo her/his main procedure or intervention; however, during the stay, s/he very often needs a set of additional procedures, carried out by diferent wards, such as diagnostic exams or specialistic consultations, which are referred to as ”outpatient services provided during hospitalization” (OSs henceforth). The admission ward has to request an OS, and wait until the external resources are available for its completion, possibly experiencing a delay.
In this paper, we aimed at verifying whether OSs have a major impact on
. To this end, we
collected the data of more than 7000 patients admitted to Azienda Ospedaliera SS. Antonio e Biagio e
transformed such data into an event log [
days. The 20-day cutof for defining ”short” versus ”long”, was not arbitrary but was set in consultation with domain experts: they determined that, based on historical patterns and clinical relevance. This expert‐driven threshold ensures the model’s outputs align with real-world expectations and decision‐making.
Classification was carried out resorting to a Long Short-Term Memory (LSTM) architecture. https://upobook.uniupo.it/giorgio.leonardi (G. Leonardi); https://upobook.uniupo.it/stefania.montani (S. Montani);
CEUR
ceur-ws.org
The very high quality of our classification results suggests that OSs play a significant role in determining , and that a better organization in the request and provision of such external services may have a positive efect on patient management and healthcare costs.
Medical process traces constitute a very important source of information, that has been adopted to
support diferent tasks within the field of process mining [
The approaches in [10, 11], instead, rely on a Transformer, an architecture that substitutes the recurrence by the attention mechanism [12]. We plan to test transformers for trace classification in our future work.
The deep learning architecture we have adopted exploits a Long Short-Term Memory (LSTM) network, as shown in Figure 1. The input consists of a trace, i.e., a sequence of activities (namely, a sequence of OSs). Each OS is preliminarily converted into a numerical format using one-hot encoding.
The architecture features two LSTM layers that capture temporal dependencies within the trace. The LSTM’s hyperparameters were chosen experimentally via small-scale tuning rather than a formal search. In particular, first LSTM layer has 256 units with a tanh activation function. It is followed by the second LSTM layer that consists of 128 units with tanh activation, that is then passed through a dropout layer with a 0.5 rate, which helps prevent overfitting by randomly deactivating half of the neurons during training. After the LSTM layers, a dense layer with 16 units and a relu activation function is applied to learn complex relationships and reduce dimensionality. Finally, a classification layer with a sigmoid activation function outputs a probability score for binary classification.
Our real world dataset encompasses a total of 7393 patients, representing a varied population with broad clinical profiles. The mean is 11.10 days, with a standard deviation of 9.01 days, and 12 OS activities on average (ranging between 2 and 236). The dataset is imbalanced, with 89.06% of the traces categorized as class 0 ( < 20 days) and 10.94% classified as class 1 ( ≥ 20 days). We then implemented an oversampling technique relying on SMOTE [13], augmenting the underrepresented class up to 2427 items.
The results are shown in Table 1. The average plots of loss and accuracy for each epoch are shown in Figure 2. The experimental outcomes describe a robust performance of the LSTM-based architecture. The precision of 0.91 and the recall of 0.90 in the short group (< 20 days – class 0) show the network’s ability to capture temporal dependencies relevant to shorter hospital stays. A precision of 0.75 and a recall of 0.78 for the long group (≥ 20 days – class 1) underscore sound performance too, although there remains room for improvement.
activation = tanh LSTM 128 Dropout 0.5
activation = tanh Dense 16 activation = relu
… activation = sigmoid Classification Layer
Output Class 0/1
Overall, classification results are good, achieving a notable accuracy of 0.87 across 30 epochs with a loss value of 0.33. Such positive outcomes testify that OSs do play a major role in determining , as OS traces classify very well according to the dimension.
In this paper, we have investigated whether OSs play a significant role in determining patient . To
this end, we have collected a large real-world event log, where each trace is the sequence of the OSs
applied to the patient at hand during her/his hospital stay. The very good classification results of these
traces along the dimension of (≥ 20 days versus < 20 days) prove that the sequence of the OSs is
strongly related to the stay duration, and should be the object of proper organizational optimization
strategies. In the future, we will therefore further analyze the OS traces, e.g., by means of process
mining techniques [
As regards the classification model, we have currently exploited a LSTM architecture; in the future, we would like to test a Transformer-based network, which is progressively becoming the state of the art also in process classification and prediction. Moreover, given the intrinsic ”black-box” nature of deep learning approaches, we plan to adopt proper explainability techniques, possibly along the lines we investigated in [14].
Acknowledgements This work was supported by the project PNRR-NODES Bando per Imprese - 20232025 - TOT-AL: Ottimizzazione della Transizione Ospedale Territorio presso l’Ospedale di Alessandria. We are grateful to R. Bellini and C. Zanelli for their work on data provision.
During the preparation of this work, the author(s) used ChatGPT, Grammarly in order to: Grammar and spelling check, Paraphrase and reword. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content. Neural Networks, Springer International Publishing, 2017, p. 477–492. URL: http://dx.doi.org/10. 1007/978-3-319-59536-8_30. doi:10.1007/978-3-319-59536-8_30. [9] M. A. Khan, H. Le, K. Do, T. Tran, A. Ghose, K. H. Dam, R. Sindhgatta, Memory-augmented neural networks for predictive process analytics, CoRR abs/1802.00938 (2018). URL: http://arxiv.org/abs/ 1802.00938. arXiv:1802.00938. [10] Z. A. Bukhsh, A. Saeed, R. M. Dijkman, Processtransformer: Predictive business process monitoring with transformer network, 2021. URL: https://arxiv.org/abs/2104.00721. arXiv:2104.00721. [11] I. Donadello, J. Ko, F. M. Maggi, J. Mendling, F. Riva, M. Weidlich, Knowledge-driven modulation of neural networks with attention mechanism for next activity prediction, 2023. URL: https: //doi.org/10.48550/arXiv.2312.08847. doi:10.48550/ARXIV.2312.08847. arXiv:2312.08847. [12] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, I. Polosukhin, Attention is all you need, in: I. Guyon, U. von Luxburg, S. Bengio, H. M. Wallach, R. Fergus, S. V. N. Vishwanathan, R. Garnett (Eds.), Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, volume 30, 2017, pp. 5998–6008. URL: https://proceedings.neurips.cc/paper/2017/hash/ 3f5ee243547dee91fbd053c1c4a845aa-Abstract.html. [13] N. V. Chawla, K. W. Bowyer, L. O. Hall, W. P. Kegelmeyer, SMOTE: synthetic minority oversampling technique, J. Artif. Intell. Res. 16 (2002) 321–357. [14] G. Leonardi, S. Montani, M. Striani, Explainable process trace classification: An application to stroke, J. Biomed. Informatics 126 (2022) 103981. URL: https://doi.org/10.1016/j.jbi.2021.103981. doi:10.1016/J.JBI.2021.103981.