Patient pathways of tuberculosis care-seeking and treatment: an individual-level analysis of National Health Insurance data in Taiwan

Introduction Patients with tuberculosis (TB) often experience difficulties in accessing diagnosis and treatment. Patient pathway analysis identifies mismatches between TB patient care-seeking patterns and service coverage, but to date, studies have only employed cross-sectional aggregate data. Methods We developed an algorithmic approach to analyse and interpret patient-level routine data on healthcare use and to construct patients’ pathways from initial care-seeking to treatment outcome. We applied this to patients with TB in a simple random sample of one million patients’ records in the Taiwan National Health Insurance database. We analysed heterogeneity in pathway patterns, delays, service coverage and patient flows between different health system levels. Results We constructed 7255 pathways for 6258 patients. Patients most commonly initially sought care at the primary clinic level, where the capacity for diagnosing TB patients was 12%, before eventually initiating treatment at higher levels. Patient pathways are extremely heterogeneous prior to diagnosis, with the 10% most complex pathways accounting for 48% of all clinical encounters, and 55% of those pathways yet to initiate treatment after a year. Extended consideration of alternative diagnoses was more common for patients aged 65 years or older and for patients with chronic lung disease. Conclusion Our study demonstrates that longitudinal analysis of routine individual-level healthcare data can be used to generate a detailed picture of TB care-seeking pathways. This allows an understanding of several temporal aspects of care pathways, including lead times to care and the variability in patient pathways.


Number of pathways
This section assesses the number of pathways rendered by the IPPA with different TOR and TOE. The changes were measured by the change rate. Figure  1 shows that the number of pathways is negative corrected with both TOR and TOE. The changes were within 5%. Figure 2 shows that increasing TOR and TOE caused higher change rates than decreasing while the number of pathways is more sensitive to TOE. • Number of Contacts, median: the length considering how many healthcare contacts happened during the patient pathways started from initial care-seeking to treatment end.
• System Delay, median: the duration from initial care-seeking to treatment start.
• Pathway Length, median: the duration from initial care-seeking to treatment end.
These indices were summarised by median, and the changes were measured by change rate. As Figure 3 shows, the numbers of contact ranged from 16 to 25 while the values were positively correlated to TOR and TOE. Figure 4 addresses the system delays, showing the values were sensitive to both TOR and TOE. When TOR and TOE were both 120 days, the change is more than being doubled compared with 60 days. Figure 5 focuses on the pathway lengths. Figure 6 summarises the sensitivity of the length of pathways. The system delay was the most sensitive to TOR and TOE, while the total pathway length was the last. TOR and TOE had equal impacts on the number of contacts in this analysis. Increasing TOR and TOE brought more changes than decreasing across these three indices. The high sensitivity of the median system delay suggested an external validation with future interview data in the same setting. A previous study, Chen et al. [1], estimated the system delay in Taiwan was 29 days (interquartile range 5-73) with the same database and TB definition as my study.
Although their assessment did not consider interrupted evaluations and patients having chronic lung conditions, and so their estimates constitute a lower-bound for our approach. Comparing with other settings, Sreeramareddy et al. [2] summarised 52 studies, finding the system delays to TB treatments ranged from 2 to 87 days, finding that the low-income and high-income settings did not have a significant difference. However, the retrieved studies in their review showed an imbalance in that the studies with longitudinal data were conducted in specific hospitals or sub-populations, while the studies that covered the general population were cross-sectional. My study, which used longitudinal data on the general population, therefore, cannot be compared with them directly. Therefore, I suggest using a retrospective design with interviewing patients embedded in the longitudinal data. This approach can validate the lengths of patient pathways from the IPPA and highlight the difference from perspectives of patients and the health system.   • Initialised at Level A Hospital: whether the initial care-seeking of pathways were in Level A hospitals.
• Interrupted Evaluation: if the pathways experienced interrupted evaluation.
• Zero Delay: indicates if the pathways started their treatment at the day of initial care-seeking.
These indices were summarised by proportion, and the changes were measured by difference. As Figure 7 shows, Initialised at Level A Hospital were no sensitive to TOE while that and TOR were positively correlated. Figure 8 highlights higher TOR led more Interrupted Evaluation but TOE had negative influence. Zero Delay in Figure 9 shows negative correlations with TOR and TOE while the two Time-outs were equally contributed. Figure 10 summarises the sensitivity of the typology of pathways. The marginal changes were usually smaller than 5%. However, Interrupted Evaluation was very sensitive to TOR and TOE compared with the other two indices.