Study design and participants

A phenomenological qualitative abductive approached study was planned to interpret the lived experience of the physicians who treat tropical infections and febrile illness. We conducted the study to understand the physicians perspective with respect to the importance of clinical, epidemiological and demographical parameters which can be useful in diagnosis. This approach is appropriate as it enables an in-depth understanding of clinicians’ subjective experiences, reasoning processes, and diagnostic challenges beyond what can be captured through quantitative methods.

The institutional ethical approval was obtained before conducting the study. The participants were recruited through purposive sampling and also by snowball sampling [13,14]. The physicians whoever treated and having experience in treatment of tropical fever and febrile illness regardless of their level of experience across Indian clinical setting were recruited. Specifically, physicians working across different level of hospital setting such as secondary, tertiary and quaternary clinical setting practicing in urban, semi urban and rural regions, were included. The physicians who were not willing to give consent were excluded from the study. Potential participants were contacted in advance and they were informed regarding the study. A written or verbal consent was obtained prior to the qualitative interview. Interview was conducted either face to face or online using platforms such as Microsoft Teams or Google meet. Date of first recruitment was 17-Jul-24 and Date of last (latest) recruitment was 3-Apr-25. The study was conducted according to Standards for Reporting Qualitative Research (SRQR) guidelines and checklist provided in the supplementary information S1 Text.

Ethics Statement: Institutional Ethics Ethical approval for the study has been obtained from the institutional ethics committee of the Kasturba Medical College and Kasturba Hospital, Manipal (IEC number: 6/2024). Informed consent has been taken before conducting the qualitative interview and confidentiality of the participants were maintained by anonymizing the interview data prior to the reporting.

Interview guide preparation & validation

Interview protocol refinement (IPR) framework method was followed during the development of interview guide. IPR involved developing, reviewing, and refining interview questions by obtaining the feedback from the subject experts. The interview guide was developed through the iterative process guided by the literature evidence and consulting with the subject experts. A draft interview guide was then created and validated by the five subject experts using refinement guide provided in IPR framework [15]. Subject experts were asked to assess the content relevance, clarity, comprehensiveness, and appropriateness of the interview guide and feedback was obtained in the form of structured comments and suggestions. We modified the interview guide based on the relevant recommendations provided by the experts during validation. The final interview guide included the questions aligned with the study objectives and was pilot tested with 2 physicians to assess clarity, flow, and timing [15]. Minor modifications were done based on the pilot testing of the interview guide. The developed interview guide covers the domains such as Clinical experience and practice, diagnostic process and Diagnosis & challenges, Parameters in diagnosis, AI in clinical practice.

The first domain focused on the clinical experience and practice of the physician with respect to the tropical fevers which include questions related to commonly encountered diseases, frequency of occurrence and seasonal pattern observed with respect to these diseases.

The second domain, i.e., Diagnosis & challenges were with respect to physicians perspective on diagnostic dilemma, difficulties arising due to overlapping clinical symptoms, challenges encountered despite the availability of laboratory tests and factors affecting diagnostic test accuracy.

The 3rd the main was pertaining to parameters in diagnosis. Here we aim to obtain the opinion on clinical and laboratory variables that might help in development of diagnostic model and to describe how these variables influence the clinical decision making. Additionally potential confounding factor that might affect the diagnosis and that can be considered as parameters in model development were also considered.

4th remain was majorly focused on the application of AI in clinical practice. This domain focused on opinions and experiences regarding desirable features for ML best clinical decision support system, it's expected accuracy levels, perceived challenges in integrating such models into routine clinical practice and the clinical setting that might be most benefited from the implementation of such model.

The domain based summary table describing the conceptual framework and its components have been provided in the Table 1.

The validated interview guide have been provided in the supplementary information S2 Text. The Question–Analysis table which helps in mapping the questions in the interview guide with the corresponding analytical approach have been provided in the supplementary information S3 Text.

Interview process

The validated interview guide was used to conduct the indepth semi structured face to face or online based interviews with the participating physicians. The interviews were conducted by a PhD student who had received formal training in qualitative interview methods. Potential participants were contacted in advance, and they were informed regarding the purpose and procedure of the study. We also informed regarding data confidentiality, and voluntary consent to the study. Prior to the interview, informed consent was obtained verbally or in writing, and the interview was audio recorded. Verbal consent was documented through audio recording, while written consent was recorded in script form. Participants were asked about their clinical experience in diagnosis and challenges, relevant diagnostic parameters and their perspective with respect to the use of AI in clinical practice. Each interview lasted approximately around 20–30 minutes. We continued the interview until we reached data saturation. Currently, during this study process, saturation was reached after eight interviews, and two additional interviews were conducted to confirm stability of the themes. Therefore, a total of ten participants were included.

Data collection and analysis

After the face-to-face interview with the physicians, the recorded audio was transcribed verbatim. Each transcript were read multiple times to ensure accuracy and data familiarization to capture all the context of the participants experience. Reflexive thematic analysis by Braun and Clarke’s framework [16] with abductive approach was used. Hence, the involved study approach moves between data-driven insights and existing conceptual understanding in an iterative manner. Initially, open coding was performed to identify patterns emerging directly from the data. Subsequently, these codes were examined in relation to existing literature and conceptual understanding of clinical reasoning and diagnostic processes. Data collection and analysis were conducted concurrently, and interviews were continued until no new codes or themes emerged.

Two independent researchers annotated the transcript and the annotated versions were validated by a subject expert. Discrepancies were resolved through the discussion and consensus. The validated transcripts were then coded to obtain the initial coding framework iteratively and updated as new insights emerged. These codes then grouped together to develop categories, followed by developing themes and sub themes. These themes were then validated with subject experts. A total of 2–3 rounds of reviews were conducted for preparing final themes and sub themes. The data management and thematic analysis were conducted using the trial version of ATLAS.ti software [17,18]. In the final themes, sub themes add the respective quotations are provided in the supplementary information S4 Text.

Participant characteristics

A total of 10 physicians, specializing in infectious diseases were requested to take part in the qualitative interview, where we conducted one-on-one indepth semi structured interview for 20–30 minutes. Data saturation was achieved during the data collection. Out of 10 participants, 7 participants were recruited through purposive sampling and 3 participants were recruited through snowball sampling. Participants aged between 31–60 years of age with experience of about 12–30 years. The interview was focused on gathering insights regarding the importance of clinical variables, associated epidemiological and demographic factors, and the challenges, dilemmas, and barriers encountered in the diagnostic process. Additionally, participants were asked to share their perspectives on the design and potential usefulness of machine learning models in clinical decision-making.

From the thematic analysis, we have obtained 4 themes with 3–5 subthemes per each theme. One focused on clinical experience & practice, 2^nd on diagnosis & challenges, 3^rd on parameters in diagnosis, and 4^th on AI in clinical practice and under every theme we obtained 5–6 subthemes. Thematic mapping of the themes and subthemes have been provided in Fig 1. Throughout these themes, implementation-associated measures such as acceptability, perceived usefulness, potential barriers, and workflow integration of AI-based solutions are embedded and highlighted as appropriate.

Theme 1: Clinical experience and practice: Clinical experience of the physicians is one of the important context which guide us to the tropical diseases they are treating, what is their experience in the clinical setting regarding the diagnosis and what are the challenges in their practices. This involves 4 subthemes as follows.

• Sub-theme 1.1: Reasons for selecting Infectious Diseases specialization-

A predominant reason was the dynamic and evolving nature of the field, with frequent emergence of new infections and a broad spectrum of clinical presentations. Unlike certain other chronic conditions, many infectious diseases are curable, and physicians found it rewarding to witness significant clinical improvement and enhanced quality of life in patients following appropriate treatment.

“There is always newer diseases and your patients become absolutely well once you treat them well, so. Those sort of push me into it”. [P1]

“It is something that is preventable and treatable, so it still gives you hope to life. So, unlike you know, there are several diseases which, wherein you can't do much and you just, you know, patients… Patients live, but then they don't, they may not live very well, post treatment. Wherein infection, if you treat right, most of the majority of them, at least people can have a good quality of life, once they come out of the disease.” [P9]

Physicians also highlighted the high burden of tropical diseases and new learnings, which created a clinical demand and professional focus toward managing these conditions.

“So since, ah, the demand for focused, Hospital acquired infections were high. I slowly, you know, I think it is like a demand driven, ah,... process which enabled me to take- take infectious diseases.” [P6]

• Sub-theme 1.2: Frequently encountered diseases in clinical practice:

All the physicians (10/10) reported that infectious disease patterns varied by season and region, with weekly case numbers ranging from about 3–110—showing that while cases increased sharply during certain seasons, some illnesses were seen all the year. Dengue fever was the most common tropical illness, followed by scrub typhus (especially in southern India), leptospirosis during the monsoon, and enteric fever and chikungunya during monsoon outbreaks. Malaria remained significant in areas like Cuttack and Mangalore but was less common in cities. Tuberculosis (including extrapulmonary forms) was a frequent cause of prolonged fever, while influenza (H1N1 and other strains), melioidosis, Kyasanur Forest Disease (KFD), and other rickettsial and Salmonella infections were also reported in various regions

“We would see dengue very commonly, scrub typhus and we would see some typhoid, chikungunya and rarely we see malaria. So, this would be the commonest acute febrile illness” [P1]

“Frequently we see dengue, malaria, leptospirosis, melioidosis. scrub typhus. Very occasionally Kyasanur forest disease. These are the common ones that we see” [P8]

“So we get about six of them. So most common being scrub typhus and leptospirosis. We also get dengue and malaria. And now we are getting more of KFD's and salmonella, sometimes if they come as a acute febrile illness.” [P5]

“That depends on season. Sometimes I can see 30 also (Smiling). Depending upon the season ma… (Smiling). This, uh.. today… marginally today I haven’t seen anything which is tropical.” [P9]

“You can say quite an outpatient wise quite a big number actually, at least 110 patients I will see in a day with either dengue or. Dengue, or influenza, like illness or acute diarrheal disease. I see at least 10 patients per day.” [P7]

• Sub-theme 1.3: Seasonal pattern of illness:

Physicians consistently reported that tropical diseases show clear seasonal trends and this has been pictorially depicted in Fig 2:

• Dengue fever and scrub typhus peak during the post-monsoon months (August–October), with dengue showing 2–3 peaks annually.
• Leptospirosis occurs mainly in the monsoon season, while typhoid and gastroenteritis rise at the end of summer due to poor water quality and hygiene.
• Mosquito-borne diseases (malaria, dengue) increase when rainfall decreases, as heavy rain suppresses breeding.
• Some physicians reported an extended disease transmission period, typically from May to December. In years with prolonged or continuous rainfall, dengue cases were observed to persist throughout the year.
• Chikungunya and viral hepatitis are common post-monsoon. Travel-related illnesses (enteric fever, malaria) often seen in travelers from northern India.
• During the dry season (December–April), scrub typhus and Kyasanur Forest Disease (KFD) rise due to tick exposure in forested areas.

“So dengue, there is a pattern. We see dengue during about two or three spikes in a year. We see typhoid also. In sort of outbreak sort of situation that is also almost many times during couple of months, rainy season especially. Leptospirosis also after the rains and during the rainy season, so that also is seasonal. Scrub typhus is the endemic. There is actually not much seasonality noted for Scrub typhus.” [P2]

“Usually depends on the rain pattern and… usually at the end of May once the rain starts or now rain has come in another two weeks, if no rain we will start having dengue. Actually, May to December is the common pattern for most of them and, yeah… peak season is that… and I think 2 years back we had continuous rain throughout the year, We had throughout the year also. So it will change based on the rain pattern most of the times.” [P10]

Theme 2: Diagnosis and challenges: The challenges with respect to the diagnosis of tropical fevers were discussed with the physician in the clinical setting. Overlapping clinical symptoms and reliability of initial presentations and laboratory parameters, are the major challenges for the differential diagnosis of tropical fevers in the clinics. The diagnosis might also depend on the epidemiological risk factors and other factors.

• Sub-theme 2.1: Challenges with respect to current diagnostics:

Physicians highlighted that although specific diagnostic tools developed for several tropical diseases, however, their high cost, prolonged turnaround time, and limited accessibility significantly reduce their real-time clinical value. The overlapping clinical features of tropical fevers further complicate differentiation, often forcing clinicians to depend on empirical or syndromic approaches rather than confirmed diagnoses.

Serological tests were reported to be nonspecific and delayed wit respect to positive reporting, as in some diseases antibodies appear only after several days of illness, resulting in missed opportunities for early treatment. Additionally, sensitivity and specificity issues across commonly used tests lead to false positives and negatives, lowering diagnostic reliability. Even gold-standard diagnostic methods, such as blood culture and polymerase chain reaction (PCR), are limited by issues of non-availability in certain clinical settings, inherent methodological constraints, and have prolonged turnaround times, which can delay diagnosis and have poor patient outcome.

These challenges are even more exacerbated in the overburdened healthcare systems of endemic regions which struggle to meet infectious disease diagnostic demand. Although molecular tests and rapid diagnostic tests (RDTs) have improved with respect to accessibility, they are still either too costly, unevenly distributed, or unreliable due to inconsistent quality. As there is diagnostic variability exist with changes in the stage of illness, many cases might be missed or diagnosed late (Time-dependent diagnostic utility of available assays).

Physicians also emphasized the lack of multiplex or differential diagnostic tools capable of concurrently identifying multiple infections, and also the influence of viral mutations which modify the clinical presentations and reduces the applicability of existing tests.

“The present available diagnostics, there are certain gaps because the serologies are nonspecific. They only after a couple of days or a week, sometimes it like for example Leptospirosis serology would be positive, maybe even after a week or so, sometime diagnosis may be delayed.” [P2]

“Many of them don't have diagnosis in the ahh.. first week, for example, leptospirosis. Antibody become positive only after five or seven days, by which day the window for appropriate treatment is over” [P3]

“Same with case for even scrub typhus. All serology become positive after about a week only. But the best time to treat is in the first one week.” [P3]

“There are certain organisms like leptospira, where neither they will grow in the culture nor we have any PCR tests readily available. So we have to depend on the antibody test. Then we missed the diagnosis in the first week unless we have a strong suspicion.” [P7]

“Chikungunya were never presented with tip of nose discoloration earlier. So now it is doing it. So these viruses undergo something called shifts and drifts.” [P9]

• Sub-theme 2.2: Challenging diseases in clinics:

Clinicians perceive several infections including leptospirosis, scrub typhus, and enteric fever as diagnostically challenging due to their non-specific signs-symptoms and delayed or unreliable test results. As per the doctors, diagnosis of leptospirosis is the most challenging since it mimics other diseases. Its conventional diagnostic test turns positive only after the first week and PCR testing is limited to mainly to tertiary centers. Scrub typhus also presents similar challenge, especially when the characteristic ‘Eschar’ is absent. The antibody tests, does not turn positive until antibodies are produced against the causative and are unreliable in the early cycle. Enteric fevers are also difficult to diagnose, as the Widal test and blood cultures lacks sensitivity and slow, especially in rural areas. Other infections including malaria, dengue, KFD, tuberculosis, and melioidosis exhibits variable degrees of diagnostic challenges. It is easy to diagnose dengue due to availability of rapid antigen tests, whereas non-falciparum malaria, extrapulmonary TB, and melioidosis are more complicated. Overall, physicians highlighted limited diagnostic infrastructure, especially in lower level of healthcare system, significantly contributing to missed or delayed diagnoses.

“So I think scrub typhus can be, if there is no eschar can be difficult diagnose. Dengue usually has a rash. So, it's reasonably simpler. But Typhoid can be quite difficult to diagnose because it doesn't have any unique standout features” [P1]

“probably scrub and typhoid? That is where we struggle. Typhoid, especially thing takes culture. It takes time, and none of the other tests are kind of conclusive. Malaria. It's OK, I think. If you are using antigen, reasonably OK. But if it is non falciparum-non vivax, then it becomes a problem. Dengue, usually we don't have that kind of challenge.” [P10]

“Sir. Dengue and KFD are more challenging, so then leptospirosis and Scrub typhus. Scrub typhus will usually have a predominant hallmark of that eschar, which you usually find and leptospirosis we know based on the clinical profile that they look very stable, but they have organ failures and subclinical hemorrhage will be there, so you should diagnosed clinically. For dengue and KFD are difficult to diagnose clinically because there are a lot of overlapping features.” [P5]

• Sub theme 2.3: Overlapping clinical features:

Physicians stressed that many tropical fevers share overlapping symptoms such as high-grade fever, headache, myalgia, malaise, fatigue, nausea, loss of appetite, and rashes, which makes early diagnosis more challenging. In the first week, most cases present as undifferentiated fever without organ-specific signs, complicating differentiation based on history alone. Common laboratory findings including thrombocytopenia, leukopenia, elevated liver enzymes, and renal impairment, further overlap across diseases. Although some clinical clues (e.g., eschar in scrub typhus, jaundice in leptospirosis, rash in dengue, joint pain in chikungunya, periodic fever in malaria, and hemorrhagic signs in KFD) may suggest specific infections, these features are often inconsistent and tend to appear only as the illness progresses.

Some of the clinical or laboratory clues were considered suggestive:

• Scrub typhus: altered sensorium, presence of eschar, hepatosplenomegaly
• Leptospirosis: severe myalgia, jaundice without oliguric renal failure, conjunctival hemorrhage
• Dengue: blanching erythematous rash, leukopenia, capillary leak in severe cases
• Chikungunya: small joint arthritis, discoloration of nose tip
• Malaria: periodic fever with chills and sweats
• KFD: prolonged APTT with normal PT, high CPK, hemorrhagic features

These features were not consistently present and often emerged only after disease progression or in ICU settings.

At least many symptoms are common for almost many of the tropical diseases, including dengue, lepto, chikungunya, malaria, enteric. So all these have,um,many overlapping symptoms. [P6]

I think fever is common. Then thrombocytopenia, that is common. Ah… then going forward, if they are sick, they have liver and renal dysfunction. And… that is common what we see. [P10]

OK, fever is a common symptom for anything. In characteristics of the fever, say for example, if it is malaria, there will be high grade fever preceded by chills and triggers and fever, followed by profuse sweating. And there can be patient, can be almost normal in between the episodes of fever, or that can be fever, one episode a day or once in 2 days. Once in three days, these are the typical manifestations, so otherwise common manifestations.[P7]

• Sub theme 2.4: Reliability of preliminary clinical indicators for early-stage diagnosis: Physicians highlighted that early diagnosis of tropical febrile illnesses depends on recognizing clinical patterns rather than single symptoms. Preliminary indicators such as fever type, rash, gastrointestinal and mucosal involvement, jaundice, conjunctival congestion, and lymphadenopathy, along with factors like travel history and illness progression, can raise diagnostic suspicion. Thrombocytopenia, leukopenia, elevated hematocrit are few basic investigations, considered helpful for screening.

However, few doctors disagreed to the above context saying that early symptoms are non specific and specific features like rashes will appear after several days of infection making the differential diagnosis difficult. Initial laboratory parameters might be pointing towards a specific disease, however they cannot be considered as confirmatory results. Additionally, many scoring systems have been developed and published in the infectious disease domain, yet their practical utility and real world applicability is limited.

So clinical features can be the like I mentioned, mainly six of them. Fever, rash, then involvement of, you know, mucosal areas and GI symptoms jaundice, conjunctival involvement, lymphadenopathy and rhabdomyolysis.If you take these clinical parameters 7–8 of them, that can be reliable in the first week only, where the antibody will not come positive. [P5]

Developing a score is quite easy, but then to translate into clinical practices where the true challenge lies, yeah. [P1]

• Sub theme 2.5: Epidemiological risk factors:

Occupation, travel history, and exposure to vectors are some of the epidemiological and environmental risk factors are few of the crucial components influencing the of clinical diagnosis. Specifically, during the early stages of disease cycle when laboratory confirmation are not available. Doctors usually feel confident to use this contextual clues to narrow the diagnostic pattern. They stated that, forest exposure is suggestive of KFD or tick-borne illness. The region-specific nature of these factors and predictive algorithms poses a major challenge to developing universally applicable diagnostic tools. Environmental variability, behavioral diversity, and occupation-based exposure make it difficult to create standardized diagnostic models that work across regions.

“Epidemiological factors means the season (Ah..) the nearby, the patients, habits and where he is traveling,the travel history. Such things are very important.” [P4]

There are a lot of people who work at night. Those who work with call centers and international offices, right? So they work during night time. And also farmers who work in fields they are more likely to be exposed to these mosquitoes during daytime and again coming to farmers, some people who walk or work in running water or stagnant water are at risk of leptospirosis, right. [P8]

Theme 3: Parameter in diagnosis:

Physicians described that diagnostic reasoning in tropical febrile illnesses is shaped by a combination of clinical, epidemiological, and laboratory parameters. In the absence of definitive early tests, clinicians rely on structured clinical judgment grounded in local disease ecology and patient presentation.

• Subtheme 3.1: Integrated Clinical Decision-Making

In routine practice, physicians adopt a syndromic approach to manage non-specific febrile illnesses, guided by geography, exposure, and symptom clusters. Physicians apply integrative frameworks (e.g., Modified Faine’s Criteria for leptospirosis) combining epidemiology, symptoms, and labs. Endemic patterns inform suspicion, for example, scrub typhus in dry livestock regions, KFD in forest zones, leptospirosis in waterlogged agricultural areas, and malaria in coastal regions. Exposure to cattle, forests, or stagnant water further refines diagnostic probability. Laboratory trends such as leukopenia (dengue), normal or elevated WBC (scrub typhus, leptospirosis), eosinopenia (typhoid, leptospirosis), and CRP levels help distinguish viral from bacterial etiologies. Clinicians combine these clinical and demographic cues to iteratively narrow differential diagnoses. Clinical decision making rules provided by physicians have been provided in supplementary information S5 Text.

“For leptospirosis we have a very defined published criteria called Modified Faine’s Criteria, which includes all what you described epidemiological thing, the clinical factors, lab parameters etc, etc and which helps us to (Ahh..) understand or the possibility of leptospirosis or try to define leptospirosis using a clinical and basic lab criteria. So for leptospirosis there is something like that. Other than that, for dengue also we have. a set of standard or most common presentations. As I mentioning, headache, retro-orbital pain and leukopenia then later lymphocytosis and thrombocytopenia. So for dengue also we have a classical picture, blood picture and symptoms. For Scrub and typhoid. (Ahh..) Really, there is nothing much, very, very specific. Or very something that would actually stand out and say that this could be scrub.” [P2]

“So we follow based on the demographic profile. So which area they come from? Like if they come from Davanagere, Honnali or unable that area which are more dry areas, they are more likely to have scrub typhus because that is the cattle, which is very common there. ” [P5]

• Subtheme 3.2: Diagnostic Testing and Emerging Alternatives

Specific microbiological tests are often limited by timing, sensitivity, and accessibility. In the early stage of illness, blood tests and PCR may not show clear results, and even culture tests do not always detect the infection reliably. Physicians emphasized the lack of affordable, standardized diagnostics outside tertiary centers. They expressed strong interest in multiplex point-of-care (POC) tools capable of detecting multiple tropical pathogens relevant to the Indian context and anticipated that AI-based platforms integrating clinical and epidemiological data could assist in early diagnostic triage, serving as supplements, not replacements, to clinical judgment.

“Then it comes to the lab diagnosis right after the clinical syndrome and the the basic labs.Then we come on to the specific microbiological test” [P2]

Other methods,(Ahh..) BioFire is a method, that is available in Microbiology department in Manipal, but it does not cover all the, tropical illnesses” [P8]

• Subtheme 3.3: Data Capture and Documentation Challenges

Clinicians reported inconsistencies in data recording. While electronic Health Information Systems capture laboratory variables, clinical histories and progress notes are often handwritten, hindering data retrieval. These are the handwritten records which are difficult for reading due to scribbled nature and usually paper-based documents, make it challenging to store, retrieve, and use it effectively. New symptoms occurring during hospitalization may go undocumented. They also acknowledged the subjective clinical interpretation, where clinical management depends on clinician’s speciality & perspective and mentioned it as “Rashomon effect”. Additionally, physicians first priority is patient management rather than recording the data.

I'm working in a facility where the -the, there is HIS, that is Health information system electronic, capturing of the health information system [P6]

Depends in the sense the approach depends on what you are suspecting clinically. OK, to give you an example, there is something called Rashomon Effect [P7]

• Subtheme 3.4: Temporal Dynamics in Diagnosis

The ‘timeline of illness’ was acknowledged as a critical parameter of diagnosis. Disease clinical presentations and test results changes with time. They mentioned that platelet counts drops after the fifth day incase of dengue; in enteric fever week-wise progression with respect to growth of causatives, and with respect to leptospirosis or scrub typhus, complications usually elevated in the second week. Choice of test mainly depends on duration of illness, with antigen or PCR-based assays preferred early and antibody-based tests later. Physicians stressed that diagnostic models, including AI tools, must incorporate this time-sensitive feature of the disease to reflect real-world clinical scenarios. [Quote 30 & 31]

Maybe normal in dengue in the initial 2–3 days, but suddenly dropping out of five days. [P3]

Whether it is in the early phase of the illness or… medium term, or it is, if it is after one week again, the picture is going to change. [P10]

Theme 4: AI in clinical practice: The physicians discussed the current use and awareness of AI use in clinical practice. The key aspects to the implementation strategy including acceptability, perceived usefulness, barriers, and workflow integration have also been reviewed. They also discussed regarding their expectations with respect to the development of AI tools and their applicability to the clinical context. They also highlighted the need for transparency in the result provided by the model. With respect to the validation, implementation and integration of the AI tools into the clinical practice, there's still a gap exists between the evidence and the real world practice.

• Subtheme 4.1: Current Use and Awareness of AI in Clinical Practice (Acceptability of AI tools)

Most physicians stated limited or no experience with conventional AI diagnostic tools and labelled themselves as “AI illiterate.” They articulated the need for structured training before profound adoption. A few mentioned using platforms like ChatGPT for clinical decision support system, though there was ambiguity about whether systems like “Track Care” be suitable as AI.

“I ah… know. I mean, I don't know if my ‘track care’ is AI with that too, it is, right? Something that we see there, this thing, that, is also kind of a AI. So that's what we have used” [P9]

“I have not used. I'm not that tech savvy to use those things, but a few of our friends they have used. They say if you give a clinical picture to our chart ChatGPT or Meta, we get differential diagnosis that that really helps.” [P7]

• Subtheme 4.2: Clinical Applicability and Expectations (Perceived usefulness)

Clinicians perceive AI as potentially effective for differential diagnosis of tropical fevers at all healthcare levels particularly in resource-limited and high-patient-load clinical care setting. They believed AI could aid in early diagnosis, guide test prioritization, and improve diagnostic efficiency in tertiary as well as quaternary care.

Physicians reported the overall diagnostic accuracy required to be above 85–95% and specificity-sensitivity more than 90–95%. They mentioned that current clinical expertise and serological methods already exceeds 80–90% and hence greater accuracy is anticipated than what is currently available. AI is seen as a supportive aid rather than a replacement for human judgment, with validation across centers considered to be essential.

“AI outcome should be more specific and better than physician himself since we are using large data. And when they can only diagnose perfectly 80-90%, AI should be more than that” [P2]

“Yeah, see. It may be it may sound a little paradoxical. It will be most beneficial in places where there is no infrastructure to evaluate the patients in a conventional way. Because say for example in a place like Bangalore where I work. We have access to all investigations. And most of the patients are either covered with insurance or affordable to get those tests done to conclude the diagnosis. In a smaller cities, there are two cities or smaller places where there is no lab facility. Patient may or may not be affordable also. I think a lot of time that treatment is based on the clinical diagnosis rather than the laboratory conclusions. If the accuracy level, if it exceeds the clinician's predictions, accuracy level of diagnosis by a tool exceeds the clinician's diagnosis. Then I think it will make a difference.” [P7]

• Subtheme 4.3: Model Development and Feature Selection (Design expectations)

Clinicians highlighted the necessity for comprehensive data capture across disease stages, including demographics, epidemiologic exposure, symptoms, lab results, therapy history, and outcomes.

They stressed the importance of local epidemiology and its reflection in AI, provide interpretable predictions, offer differential diagnoses, and exhibit lab trends over time. Explainability (“no black box”) was viewed as critical for clinical acceptance, with tools like SHAP or LIME cited as useful for transparency.

“So you need to collect data regarding everyone of their symptoms, duration of symptoms, nature of those symptoms. Right. Let's say for example if you talk about fever. So in fever in the history, there are so many questions that we ask right, on set of duration, severity, you know diagonal variation. As well as you know what it responds to, what makes it worse, what time of the day all those will be there. So each of these have to be captured.” [P8]

“Need to open the black box” [P3]

• Subtheme 4.4: Validation, Integration, and Implementation Challenges (Barriers and feasibility)

Physicians empathised the need for real-world validation to be qualified for trusted clinical use of AI. For the reason that, regional disease patterns, patient heterogeneity, and incomplete clinical histories the few factors that could influence the validation. They also stressed, seamless integration of such tools with electronic medical records (EMRs) and laboratory databases, which could reduce works such as manual data entry and help in improved workflow efficiency. However, this integration necessitates a strong information technology (IT) infrastructure including secure servers, and dedicated maintenance support. In addition, healthcare professionals needs continuous training to use the updated technologies.

Ethical considerations also play a critical role, including ensuring data privacy, informed consent, and appropriate anonymization. Transparent and responsible use of AI systems can enhance patient confidence, improve adherence, and increase the trust in healthcare institutions.

“The challenge will be only when, -- once it is proven to be useful, I don't find much of a problem. Except that we need a good IT. For example, if you have to be implemented. The main challenge is to validate it.” [P3]

“Yeah. For example, who will enter all these variables into the.. That's the challenge. So, ah, if a particular health system, for example, ah, anywhere the labs are coming in the electronic medical record only. So the electronic, the AI has to be integrated to the current electronic-electronic medical records of the hospital. So that you know it will be an, it has to be integrated to the current electronic health record system, then only it will become useful. For example, if it is just an app, then who is going to populate this app? Nobody has time, right? So the challenges integrating the AI to the existing healthcare system” [P6]

• Subtheme 4.5: Data Quality and Model Validation Requirements (Trust and reliability requirements)

Trust in AI tools depends on accurate data set that include clinical, epidemiologic, and lab data, used for tool development and prospective validation of the tool. Clinicians expect that, external validation needs to be ideally conducted within six hours of admission, to be the ideal decision support system and provide timely decision especially in critical care.

“So, I would first think that, how they investigators thought to their process very carefully. And I would really think what they are allowed to enter into the algorithm. So how they set up a really good data set. First prospectively, to enter the algorithm and train the algorithm, if that is very patchy, I would not even use the algorithm.”[P1]

“Validation is the one challenge. Internal validation then external validation. If all that is done, then it's just question of having proper servers and um… getting ah… the… variable fed in and then business model. How we want to take it forward.” [P3]

Overall, physicians’ perspectives across themes suggest that while AI-based tools are perceived as potentially useful for supporting differential diagnosis, their adoption in clinical practice depends on multiple factors. These include trust in model outputs, integration within existing workflows, availability of high-quality data, and context-specific applicability. Key barriers identified include limitations in diagnostic infrastructure, variability in clinical presentation, and challenges in data capture. These findings highlight the importance of aligning AI tool development with real-world clinical needs and constraints.