The "Left without being seen" (LWBS) rate describes the percentage of patients who leave Emergency Departments (ED) after triage without being seen by a doctor. This rate reflects ED operational efficiency because a high LWBS rate is typically related to long waiting times and overcrowding. The LWBS phenomenon can negatively impact the quality of healthcare, as patients who leave the ED do not receive the necessary medical attention, so their health condition could worsen, leading to higher costs for the National Health Service. While descriptive analytics is commonly employed to address this phenomenon, little attention is given to predictive and prescriptive analytics. Previous Machine Learning (ML) studies proposed predictive methods capable of identifying categories of ED patients at risk of leaving, without capturing their behavior over time. At the same time, Operations Research (OR) studies focus on other metrics, such as the Door To Doctor Time (DTDT) and the ED Length of Stay (EDLOS). In this thesis, ML methods are employed to perform a Survival Analysis, estimating the probability of each patient leaving the ED before being seen, according to their characteristics, as the waiting time increases. The work proposes a MILP based approach, informed by predictions from Survival models, integrated with a Discrete Event Simulation (DES) model that mimics the ED dynamics to minimize the percentage of LWBS, DTDT and EDLOS, investigating the trade off between these indices. In order to evaluate the effectiveness of the proposed methods, a prioritization algorithm was implemented to replicate a policy similar to the real case and compared with the MILP based approach. ML methods were compared through a quantitative analysis in terms of accuracy and impact of their use on optimization of visit scheduling. Keywords: Emergency Department, Left without being seen, Survival Analysis, Machine Learning, Mixed Integer Linear Programming.
Il tasso di "Left without being seen" (LWBS) descrive la percentuale di pazienti che lasciano i pronto soccorso (PS) dopo il triage, senza essere stati visitati da un medico. Questo tasso riflette l’efficienza operativa del PS, perché un alto tasso di LWBS è tipicamente correlato a un lungo tempo di attesa e al sovraffollamento. Il fenomeno del LWBS può impattare negativamente la qualità delle cure, dal momento che i pazienti che lasciano il PS non ricevono la necessaria attenzione medica, quindi la loro condizione di salute potrebbe peggiorare, portando a costi maggiori per il Sistema Sanitario Nazionale. Mentre l’analisi descrittiva è spesso utilizzata per trattare questo fenomeno, viene data poca attenzione all’analisi predittiva e prescrittiva. Studi precedenti di Machine Learning (ML) hanno proposto metodi predittivi in grado di identificare categorie di pazienti PS a rischio di abbandono, senza descrivere il loro comportamento nel corso del tempo. Al tempo stesso studi di Ricerca Operativa (Operations Research, OR) si focalizzano su altri indici, come il Door-To-Doctor Time (DTDT) e la permanenza totale in pronto soccorso (ED Length of Stay, EDLOS). In questa tesi vengono impiegati metodi ML per svolgere un’analisi di sopravvivenza, stimando la probabilità, per ogni paziente, di lasciare il PS prima di essere visitato, secondo le sue caratteristiche, al crescere del tempo. Il lavoro propone un modello di Programmazione Lineare Intera Mista (Mixed-Integer Linear Programming, MILP), informato dalle predizioni provenienti dai modelli di sopravvivenza, integrato con un modello di Simulazione a Eventi Discreti (Discrete Event Simulation, DES) che simula le dinamiche del PS al fine di minimizzare la percentuale di LWBS, il DTDT e l’EDLOS, studiando il trade-off tra questi indici. Al fine di valutare l’efficacia del metodo proposto, è stato implementato un algoritmo di prioritizzazione che replicasse una politica simile al caso reale ed è stato paragonato all’approccio basato sulla MILP. I metodi ML sono stati comparati attraverso un’analisi quantitativa in termini di accuratezza e impatto del loro utilizzo sull’ottimizzazione della programmazione delle visite. Parole chiave: Pronto Soccorso, Left without being seen, Analisi di Sopravvivenza, Machine Learning, Programmazione Lineare Intera Mista.
Pazienti che abbandonano il pronto soccorso senza essere visitati: un approccio di ottimizzazione basato sull'analisi di sopravvivenza
MEINI, VITTORIO
2023/2024
Abstract
The "Left without being seen" (LWBS) rate describes the percentage of patients who leave Emergency Departments (ED) after triage without being seen by a doctor. This rate reflects ED operational efficiency because a high LWBS rate is typically related to long waiting times and overcrowding. The LWBS phenomenon can negatively impact the quality of healthcare, as patients who leave the ED do not receive the necessary medical attention, so their health condition could worsen, leading to higher costs for the National Health Service. While descriptive analytics is commonly employed to address this phenomenon, little attention is given to predictive and prescriptive analytics. Previous Machine Learning (ML) studies proposed predictive methods capable of identifying categories of ED patients at risk of leaving, without capturing their behavior over time. At the same time, Operations Research (OR) studies focus on other metrics, such as the Door To Doctor Time (DTDT) and the ED Length of Stay (EDLOS). In this thesis, ML methods are employed to perform a Survival Analysis, estimating the probability of each patient leaving the ED before being seen, according to their characteristics, as the waiting time increases. The work proposes a MILP based approach, informed by predictions from Survival models, integrated with a Discrete Event Simulation (DES) model that mimics the ED dynamics to minimize the percentage of LWBS, DTDT and EDLOS, investigating the trade off between these indices. In order to evaluate the effectiveness of the proposed methods, a prioritization algorithm was implemented to replicate a policy similar to the real case and compared with the MILP based approach. ML methods were compared through a quantitative analysis in terms of accuracy and impact of their use on optimization of visit scheduling. Keywords: Emergency Department, Left without being seen, Survival Analysis, Machine Learning, Mixed Integer Linear Programming.File | Dimensione | Formato | |
---|---|---|---|
Tesi_Magistrale_CORRETTA_A.pdf
accesso aperto
Dimensione
1.34 MB
Formato
Adobe PDF
|
1.34 MB | Adobe PDF | Visualizza/Apri |
È consentito all'utente scaricare e condividere i documenti disponibili a testo pieno in UNITESI UNIPV nel rispetto della licenza Creative Commons del tipo CC BY NC ND.
Per maggiori informazioni e per verifiche sull'eventuale disponibilità del file scrivere a: unitesi@unipv.it.
https://hdl.handle.net/20.500.14239/28435